ReportPart1.md

Time series regression - Store sales forecasting, part 1

Ahmet Zamanis

• Introduction
• Data preparation
• Overview of hybrid modeling approach
• Exploratory analysis 1: Time & calendar effects
  • Trend
  • Seasonality
  • Autocorrelation & partial autocorrelation
  • April 2016 Earthquake
• Feature engineering: Time & calendar features
  • Calendar effects & weekly seasonality features
  • Trend & monthly seasonality features
• Model 1: Time effects decomposition
  • Preprocessing
  • Model specification
  • Model validation: Predicting 2017 sales
  • Rolling crossvalidation
  • Residuals diagnosis & stationarity
  • Time decomposition
• Exploratory analysis 2: Lags & covariates
  • Covariates stationarity & differencing
  • Sales features
  • Oil features
  • Onpromotion features
  • Transactions features
• Model 2: Lags & covariates
  • Preprocessing
  • Model specification
  • Model validation: Predicting 2017 sales
  • ARIMA model summary
  • Backtesting / Historical forecasts
  • Residuals diagnosis & stationarity
• Conclusion

Introduction

This is a report on time series analysis & regression modeling, performed in Python, mainly with the Darts library. The dataset is from the Kaggle Store Sales - Time Series Forecasting competition. The data consists of daily sales data for an Ecuadorian supermarket chain between 2013 and 2017. This is part 1 of the analysis, which will focus only on forecasting the daily national sales of the chain, aggregated across all stores and categories. In part 2, we will forecast the sales in each category - store combination as required by the competition, and apply hierarchical reconciliation methods.

The main information source used extensively for this analysis is the textbook Forecasting: Principles and Practice, written by Rob J. Hyndman and George Athanasopoulos. The book is the most complete source on time series analysis & forecasting I could find. It uses R and the tidyverts libraries for its examples, which have the best time series analysis workflow among the R libraries I’ve tried.

Show code

# Import libraries
import pandas as pd 
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns

# Set printing options
np.set_printoptions(suppress=True, precision=4)
pd.options.display.float_format = '{:.4f}'.format
pd.set_option('display.max_columns', None)

# Set plotting options
plt.rcParams['figure.dpi'] = 300
plt.rcParams['savefig.dpi'] = 300
plt.rcParams["figure.autolayout"] = True

Data preparation

Show code

# Load original datasets
df_train = pd.read_csv("./OriginalData/train.csv", encoding="utf-8")
df_stores = pd.read_csv("./OriginalData/stores.csv", encoding="utf-8")
df_oil = pd.read_csv("./OriginalData/oil.csv", encoding="utf-8")
df_holidays = pd.read_csv("./OriginalData/holidays_events.csv", encoding="utf-8")
df_trans = pd.read_csv("./OriginalData/transactions.csv", encoding="utf-8")

The data is split into several .csv files. train.csv and test.csv are the main datasets, consisting of daily sales data. The training data ranges from 01-01-2013 to 15-08-2017, and the testing data consists of the following 16 days until the end of August 2017. We won’t do a competition submission in part 1, so we won’t work on the testing data.

Show code

print(df_train.head(5))

   id        date  store_nbr      family  sales  onpromotion
0   0  2013-01-01          1  AUTOMOTIVE 0.0000            0
1   1  2013-01-01          1   BABY CARE 0.0000            0
2   2  2013-01-01          1      BEAUTY 0.0000            0
3   3  2013-01-01          1   BEVERAGES 0.0000            0
4   4  2013-01-01          1       BOOKS 0.0000            0

• For each day, we have the sales in each store (out of a possible 54) and each product category (out of a possible 33). This amounts to 1782 time series that need to be forecasted for the competition.

• onpromotion is the number of items on sale in one day, in one category & store combination.

stores.csv contains more information about each store: The city, state, store type and store cluster.

Show code

print(df_stores.head(5))

   store_nbr           city                           state type  cluster
0          1          Quito                       Pichincha    D       13
1          2          Quito                       Pichincha    D       13
2          3          Quito                       Pichincha    D        8
3          4          Quito                       Pichincha    D        9
4          5  Santo Domingo  Santo Domingo de los Tsachilas    D        4

holidays.csv contains information about local (city-wide), regional (state-wide) and national holidays, and some special nation-wide events that occured in the time period. We will use these along with the stores’ location data to create calendar features, to adjust for effects of holidays and special events on sales.

Show code

print(df_holidays.head(5))

         date     type    locale locale_name                    description  \
0  2012-03-02  Holiday     Local       Manta             Fundacion de Manta   
1  2012-04-01  Holiday  Regional    Cotopaxi  Provincializacion de Cotopaxi   
2  2012-04-12  Holiday     Local      Cuenca            Fundacion de Cuenca   
3  2012-04-14  Holiday     Local    Libertad      Cantonizacion de Libertad   
4  2012-04-21  Holiday     Local    Riobamba      Cantonizacion de Riobamba   

   transferred  
0        False  
1        False  
2        False  
3        False  
4        False  

oil.csv consists of the daily oil prices in the time period. Ecuador has an oil-dependent economy, so this may be an useful predictor of the cyclicality in supermarket sales. We don’t have the oil price for the first day of the data.

Show code

print(df_oil.head(5))

         date  dcoilwtico
0  2013-01-01         NaN
1  2013-01-02     93.1400
2  2013-01-03     92.9700
3  2013-01-04     93.1200
4  2013-01-07     93.2000

transactions.csv consists of the daily number of transactions at each store. Note that only store 25 is recorded for the first day, so it’s likely all stores aren’t open in all days.

Show code

print(df_trans.head(5))

         date  store_nbr  transactions
0  2013-01-01         25           770
1  2013-01-02          1          2111
2  2013-01-02          2          2358
3  2013-01-02          3          3487
4  2013-01-02          4          1922

We will rename some columns from the datasets and merge the supplementary information into the main sales dataset. We’ll aggregate daily transactions across all stores beforehand, as we are only interested in predicting daily national sales.

Show code

# Rename columns
df_holidays = df_holidays.rename(columns = {"type":"holiday_type"})
df_oil = df_oil.rename(columns = {"dcoilwtico":"oil"})
df_stores = df_stores.rename(columns = {
  "type":"store_type", "cluster":"store_cluster"})

# Aggregate daily transactions across all stores
df_trans_agg = df_trans.groupby("date").transactions.sum()

# Add columns from oil, stores and transactions datasets into main data
df_train = df_train.merge(df_trans_agg, on = ["date"], how = "left")
df_train = df_train.merge(df_oil, on = "date", how = "left")
df_train = df_train.merge(df_stores, on = "store_nbr", how = "left")

Incorporating the holidays information into the sales dataset will require more work. We’ll create a four binary columns that record whether a day is a local, regional or national holiday, or had a special event.

Show code

# Split holidays data into local, regional, national and events
events = df_holidays[df_holidays["holiday_type"] == "Event"].copy()
df_holidays = df_holidays.drop(labels=(events.index), axis=0)
local = df_holidays.loc[df_holidays["locale"] == "Local"].copy()
regional = df_holidays.loc[df_holidays["locale"] == "Regional"].copy()
national = df_holidays.loc[df_holidays["locale"] == "National"].copy()

There are cases of multiple holidays or events sharing the same date and locale. We’ll inspect the duplicates and drop them so we don’t have issues in feature engineering.

• Rows with transferred = True are dates that are normally holidays, but the holiday was transferred to another date. In other words, these are not holidays in effect.

• Rows with holiday_type = Transfer are dates that are not normally holidays, but had another holiday transferred to this date. In other words, these are holidays in effect.

• Rows with holiday_type = Bridge are dates that are not normally holidays, but were added to extend preceding / following holidays.

Show code

# Inspect local holidays sharing same date & locale
print(local[local.duplicated(["date", "locale_name"], keep = False)])

           date holiday_type locale locale_name  \
264  2016-07-24   Additional  Local   Guayaquil   
265  2016-07-24     Transfer  Local   Guayaquil   

                         description  transferred  
264         Fundacion de Guayaquil-1        False  
265  Traslado Fundacion de Guayaquil        False  

Show code

# Drop the transfer row
local = local.drop(265, axis = 0)

Show code

# Inspect regional holidays sharing same date & locale. None exist
print(regional[regional.duplicated(["date", "locale_name"], keep = False)])

Empty DataFrame
Columns: [date, holiday_type, locale, locale_name, description, transferred]
Index: []

Show code

# Inspect national holidays sharing same date & locale
print(national[national.duplicated(["date"], keep = False)])

           date holiday_type    locale locale_name                description  \
35   2012-12-24       Bridge  National     Ecuador             Puente Navidad   
36   2012-12-24   Additional  National     Ecuador                  Navidad-1   
39   2012-12-31       Bridge  National     Ecuador  Puente Primer dia del ano   
40   2012-12-31   Additional  National     Ecuador       Primer dia del ano-1   
156  2014-12-26       Bridge  National     Ecuador             Puente Navidad   
157  2014-12-26   Additional  National     Ecuador                  Navidad+1   

     transferred  
35         False  
36         False  
39         False  
40         False  
156        False  
157        False  

Show code

# Drop bridge days
national = national.drop([35, 39, 156], axis = 0)

Show code

# Inspect events sharing same date
print(events[events.duplicated(["date"], keep = False)])

           date holiday_type    locale locale_name          description  \
244  2016-05-08        Event  National     Ecuador  Terremoto Manabi+22   
245  2016-05-08        Event  National     Ecuador      Dia de la Madre   

     transferred  
244        False  
245        False  

Show code

# Drop the earthquake row as it is a one-time event
events = events.drop(244, axis = 0)

After getting rid of duplicates, we can create binary columns that signify whether a date was a local / regional / national holiday / event. We’ll merge these back into the sales data.

Show code

# Add local_holiday binary column to local holidays data, to be merged into main 
# data.
local["local_holiday"] = (
  local.holiday_type.isin(["Transfer", "Additional", "Bridge"]) |
  ((local.holiday_type == "Holiday") & (local.transferred == False))
  ).astype(int)

# Add regional_holiday binary column to regional holidays data
regional["regional_holiday"] = (
  regional.holiday_type.isin(["Transfer", "Additional", "Bridge"]) |
  ((regional.holiday_type == "Holiday") & (regional.transferred == False))
  ).astype(int)

# Add national_holiday binary column to national holidays data
national["national_holiday"] = (
  national.holiday_type.isin(["Transfer", "Additional", "Bridge"]) |
  ((national.holiday_type == "Holiday") & (national.transferred == False))
  ).astype(int)

# Add event column to events
events["event"] = 1

# Merge local holidays binary column to main data, on date and city
local_merge = local.drop(
  labels = [
    "holiday_type", "locale", "description", "transferred"], axis = 1).rename(
      columns = {"locale_name":"city"})
df_train = df_train.merge(local_merge, how="left", on=["date", "city"])
df_train["local_holiday"] = df_train["local_holiday"].fillna(0).astype(int)

# Merge regional holidays binary column to main data
regional_merge = regional.drop(
  labels = [
    "holiday_type", "locale", "description", "transferred"], axis = 1).rename(
      columns = {"locale_name":"state"})
df_train = df_train.merge(regional_merge, how="left", on=["date", "state"])
df_train["regional_holiday"] = df_train["regional_holiday"].fillna(0).astype(int)

# Merge national holidays binary column to main data, on date
national_merge = national.drop(
  labels = [
    "holiday_type", "locale", "locale_name", "description", 
    "transferred"], axis = 1)
df_train = df_train.merge(national_merge, how="left", on="date")
df_train["national_holiday"] = df_train["national_holiday"].fillna(0).astype(int)

# Merge events binary column to main data
events_merge = events.drop(
  labels = [
    "holiday_type", "locale", "locale_name", "description", 
    "transferred"], axis = 1)
df_train = df_train.merge(events_merge, how="left", on="date")
df_train["event"] = df_train["event"].fillna(0).astype(int)

We’ll set the date column to a DateTimeIndex, and view the sales data with the added columns. Remember that the transactions column is now the national transactions across all stores in one day.

Show code

# Set datetime index
df_train = df_train.set_index(pd.to_datetime(df_train.date))
df_train = df_train.drop("date", axis=1)
print(df_train.head(5))

            id  store_nbr      family  sales  onpromotion  transactions  oil  \
date                                                                           
2013-01-01   0          1  AUTOMOTIVE 0.0000            0      770.0000  NaN   
2013-01-01   1          1   BABY CARE 0.0000            0      770.0000  NaN   
2013-01-01   2          1      BEAUTY 0.0000            0      770.0000  NaN   
2013-01-01   3          1   BEVERAGES 0.0000            0      770.0000  NaN   
2013-01-01   4          1       BOOKS 0.0000            0      770.0000  NaN   

             city      state store_type  store_cluster  local_holiday  \
date                                                                    
2013-01-01  Quito  Pichincha          D             13              0   
2013-01-01  Quito  Pichincha          D             13              0   
2013-01-01  Quito  Pichincha          D             13              0   
2013-01-01  Quito  Pichincha          D             13              0   
2013-01-01  Quito  Pichincha          D             13              0   

            regional_holiday  national_holiday  event  
date                                                   
2013-01-01                 0                 1      0  
2013-01-01                 0                 1      0  
2013-01-01                 0                 1      0  
2013-01-01                 0                 1      0  
2013-01-01                 0                 1      0  

With financial data, it’s a good idea to perform CPI adjustment before analysis, to remove the effects of inflation / deflation on our models and predictions. We’ll CPI adjust the sales and oil prices columns with 2010 as our base year. The CPI values for Ecuador in the time period were retrieved here.

• We’ll use the yearly CPI values for simplicity, but it’s possible to use monthly CPI as well.

Show code

# CPI adjust sales and oil, with CPI 2010 = 100
cpis = {
  "2010": 100, "2013": 112.8, "2014": 116.8, "2015": 121.5, "2016": 123.6, 
  "2017": 124.1
  }
  
for year in [2013, 2014, 2015, 2016, 2017]:
  df_train.loc[df_train.index.year == year, "sales"] = df_train.loc[
    df_train.index.year == year, "sales"] / cpis[str(year)] * cpis["2010"]
  df_train.loc[df_train.index.year == year, "oil"] = df_train.loc[
    df_train.index.year == year, "oil"] / cpis[str(year)] * cpis["2010"]

We have some columns with missing values in our training data.

Show code

# Check missing values in each column
pd.isnull(df_train).sum()

id                       0
store_nbr                0
family                   0
sales                    0
onpromotion              0
transactions          3564
oil                 928422
city                     0
state                    0
store_type               0
store_cluster            0
local_holiday            0
regional_holiday         0
national_holiday         0
event                    0
dtype: int64

We will interpolate the missing values in the oil and transactions columns using time interpolation.

• This performs linear interpolation, but also takes the date-time index of observations into account.

• If we don’t set limit_direction = "both", the first value for oil won’t be interpolated.

Show code

df_train["oil"] = df_train["oil"].interpolate("time", limit_direction = "both")
df_train["transactions"] = df_train["transactions"].interpolate(
  "time", limit_direction = "both")

We will now aggregate daily sales across all categories and stores, to retrieve our target variable. We have 1684 days of national sales data.

Show code

sales = df_train.groupby("date").sales.sum()
sales

date
2013-01-01     2226.6126
2013-01-02   439798.2429
2013-01-03   320444.3538
2013-01-04   314237.3024
2013-01-05   423182.7316
                 ...    
2017-08-11   665893.4102
2017-08-12   638703.0903
2017-08-13   697533.9867
2017-08-14   613152.6238
2017-08-15   614554.3400
Name: sales, Length: 1684, dtype: float64

We will create a Darts TimeSeries with our target variable.

Show code

from darts import TimeSeries

ts_sales = TimeSeries.from_series(
  sales, 
  freq="D" # Time series frequency is daily
  )
print(ts_sales)

<TimeSeries (DataArray) (date: 1688, component: 1, sample: 1)>
array([[[  2226.6126]],

       [[439798.2429]],

       [[320444.3538]],

       ...,

       [[697533.9867]],

       [[613152.6238]],

       [[614554.34  ]]])
Coordinates:
  * date       (date) datetime64[ns] 2013-01-01 2013-01-02 ... 2017-08-15
  * component  (component) object 'sales'
Dimensions without coordinates: sample
Attributes:
    static_covariates:  None
    hierarchy:          None

• Each Pandas series / dataframe column passed is stored as a component in the Darts TS. The date-time index is stored in .time_index. We had 1684 rows in our Pandas series, but the Darts TS has 1688 dates. This means our series had some missing dates, which Darts created automatically. We’ll fill in the values for these dates later.

• To create a multivariate time series, we create a Pandas dataframe with each time series as a column, and a common date-time index. When we pass this dataframe to TimeSeries, we’ll have each time series as a component. If the multivariate time series has a hierarchy, i.e. if they sum up together in a certain way, we can map that hierarchy as a dictionary to later perform hierarchical reconciliation. We will explore this further in part 2 of the analysis.

• Static covariates are time-invariant covariates that may be used as predictors in global models (models trained on multiple time series at once). They are stored together with the target series in the Darts TS. In our case, the location, type or cluster of a store may be used as static covariates, but for part 1 of our analysis we are looking at national sales, so these aren’t applicable.

Overview of hybrid modeling approach

A time series can be written as the sum of several components:

• Trend: The long-term change.

• Seasonality: A fluctuation (or several) that repeats based on a fixed, known time period, often related to how we measure time and schedule our lives. For example, the fluctuation of retail store sales across days of a week, or electricity usage across hours of a day.

• Cyclicality: A fluctuation that does not repeat based on a fixed, known time period, often caused by temporary factors. For example, the effect of a sharp increase / decrease in oil prices on car sales.

• Remainder / Error: The unpredictable component of the time series, at least with the available data and methods.

When analyzing a time series with plots, it can be difficult to determine the nature and causes of fluctuations. It can be especially tricky to tell apart the cyclical effects from repeating seasonality. Because of this, we will split our analysis and modeling into two steps:

• In step 1, we will analyze the time effects: The trend, seasonality and calendar effects (such as holidays & events). We’ll build a model that predicts these effects and remove the predictions from the time series, leaving the effects of cyclicality and the unpredictable component. This is called time decomposition.

• In step 2, we will re-analyze the time-decomposed time series, this time considering the effects of covariates and lagged values of sales itself as predictors, to try and account for the cyclicality. We’ll build a model that uses these predictors, train it on the decomposed sales, and add up the predictions of both models to arrive at our final predictions.

• This approach is called a hybrid model. It allows us to differentiate and separately model different time series components. We can also use more than one type of algorithm and combine their strengths.

Exploratory analysis 1: Time & calendar effects

Trend

Let’s start by analyzing the overall trend in sales, with a simple time series plot. Darts offers the ability to plot time series quickly.

Show code

_ =  ts_sales.plot()
_ =  plt.ylabel("Daily sales, millions")
plt.show()
plt.close()

The time series plot shows us several things:

• Supermarket sales show an increasing trend over the years. The trend is close to linear overall, but the rate of increase declines roughly from the start of 2015. Consider one straight line from 2013 to 2015, and a second, less steep one from 2015 to the end.

• Sales mostly fluctuate around the trend with a repeating pattern, which suggests strong seasonality. However, there are also sharp deviations from the trend in certain periods, mainly across 2014 and at the start of 2015. These are likely cyclical in nature.

• The “waves” of seasonal fluctuations seem to be getting bigger over time. This suggests we should use a multiplicative time decomposition instead of additive.

• Sales decline very sharply in the first day of every year, almost to zero.

Seasonality

Annual seasonality

Let’s look at annual seasonality: How sales fluctuate over a year based on quarters, months, weeks of a year and days of a year. In the plots below, we have the daily sales averaged by each respective calendar period, colored by each year in the data. The confidence bands indicate the minimum and maximum daily sales in each respective period (in the last plot, we just have the daily sales without any averaging).

Show code

# FIG1: Annual seasonality, period averages
fig1, axes1 = plt.subplots(2,2, sharey=True)
_ =  fig1.suptitle('Annual seasonality: Average daily sales in given time periods, millions')

# Average sales per quarter of year
_ =  sns.lineplot(
  ax = axes1[0,0],
  x = sales.index.quarter.astype(str), 
  y = (sales / 1000000), 
  hue = sales.index.year.astype(str), data=sales, legend=False)
_ =  axes1[0,0].set_xlabel("quarter", fontsize=8)
_ =  axes1[0,0].set_ylabel("sales", fontsize=8)
_ =  axes1[0,0].tick_params(axis='both', which='major', labelsize=6)

# Average sales per month of year
_ =  sns.lineplot(
  ax = axes1[0,1],
  x = sales.index.month.astype(str), 
  y = (sales / 1000000), 
  hue = sales.index.year.astype(str), data=sales)
_ =  axes1[0,1].set_xlabel("month", fontsize=8)
_ =  axes1[0,1].set_ylabel("sales",fontsize=8)
_ =  axes1[0,1].legend(title = "year", bbox_to_anchor=(1.05, 1.0), fontsize="small", loc='best')
_ =  axes1[0,1].tick_params(axis='both', which='major', labelsize=6)

# Average sales per week of year
_ =  sns.lineplot(
  ax = axes1[1,0],
  x = sales.index.week, 
  y = (sales / 1000000), 
  hue = sales.index.year.astype(str), data=sales, legend=False)
_ =  axes1[1,0].set_xlabel("week of year", fontsize=8)
_ =  axes1[1,0].set_ylabel("sales",fontsize=8)
_ =  axes1[1,0].tick_params(axis='both', which='major', labelsize=6)
_ =  axes1[1,0].xaxis.set_ticks(np.arange(0, 52, 10))

# Average sales per day of year
_ =  sns.lineplot(
  ax = axes1[1,1],
  x = sales.index.dayofyear, 
  y = (sales / 1000000), 
  hue = sales.index.year.astype(str), data=sales, legend=False)
_ =  axes1[1,1].set_xlabel("day of year", fontsize=8)
_ =  axes1[1,1].set_ylabel("sales",fontsize=8)
_ =  axes1[1,1].tick_params(axis='both', which='major', labelsize=6)
_ =  axes1[1,1].xaxis.set_ticks(np.arange(0, 365, 100))

# Show fig1
plt.show()
plt.close("all")

• Quarterly: Sales do not seem to have a considerable quarterly seasonality pattern. However, the plot still shows us a few things:

  • Sales generally slightly increase over a year, but the trend may be stagnating / declining at the end of our data in 2017.

  • In Q2 2014, there was a considerable drop. Sales declined almost to the level of Q2 2013. This was likely a cyclical effect.

• Monthly: Sales do seem to fluctuate slightly over months, but there’s no clear seasonal pattern that’s apparent across all years. However, sales seem to sharply increase in November and December every year, likely due to Christmas.

  • The cyclicality in 2014 is seen in more detail: Sales dropped almost to their 2013 levels in certain months, and recovered sharply in others.

  • We also see a considerable drop in the first half of 2015, where sales dropped roughly to 2014 levels, followed by a recovery.

  • There is a sharp increase in April 2016, where sales were as high as 2017 levels. This is due to a large earthquake that happened in April 16 2016, and the related relief efforts.

• Weekly: The seasonal patterns are more visible in the weekly plot, as we see the “waves” of fluctuation line up across years. It’s very likely the data has strong weekly seasonality, which is what we’d expect from supermarket sales.

  • The data for 2017 ends after August 15, so the sharp decline afterwards is misleading, though there may still be a stagnation / decline in the overall trend.

  • The sharp decline at the end of 2016 is also misleading, as 2016 was a 366-day year.

• Daily: This plot is a bit noisy, but the very similar fluctuations across all years indicate the data is strongly seasonal. It also highlights some cyclical effects such as the April 2016 earthquake and the 2014 drops.

Another way to look at annual seasonality is to average sales in a certain calendar period across all years, without grouping by year.

Show code

# FIG1.1: Annual seasonality, averaged over years
fig11, axes11 = plt.subplots(2,2, sharey=True)
_ = fig11.suptitle('Annual seasonality: Average daily sales in given time periods,\n across all years, millions');

# Average sales per quarter of year
_ = sns.lineplot(
  ax = axes11[0,0],
  x = sales.index.quarter.astype(str), 
  y = (sales / 1000000), 
  data=sales, legend=False)
_ = axes11[0,0].set_xlabel("quarter", fontsize=8)
_ = axes11[0,0].set_ylabel("sales", fontsize=8)
_ = axes11[0,0].tick_params(axis='both', which='major', labelsize=6)

# Average sales per month of year
_ = sns.lineplot(
  ax = axes11[0,1],
  x = sales.index.month.astype(str), 
  y = (sales / 1000000), 
  data=sales)
_ = axes11[0,1].set_xlabel("month", fontsize=8)
_ = axes11[0,1].set_ylabel("sales",fontsize=8)
_ = axes11[0,1].tick_params(axis='both', which='major', labelsize=6)

# Average sales per week of year
_ = sns.lineplot(
  ax = axes11[1,0],
  x = sales.index.week, 
  y = (sales / 1000000), 
  data=sales, legend=False)
_ = axes11[1,0].set_xlabel("week of year", fontsize=8)
_ = axes11[1,0].set_ylabel("sales",fontsize=8)
_ = axes11[1,0].tick_params(axis='both', which='major', labelsize=6)
_ = axes11[1,0].xaxis.set_ticks(np.arange(0, 52, 10))

# Average sales per day of year
_ = sns.lineplot(
  ax = axes11[1,1],
  x = sales.index.dayofyear, 
  y = (sales / 1000000), 
  data=sales, legend=False)
_ = axes11[1,1].set_xlabel("day of year", fontsize=8)
_ = axes11[1,1].set_ylabel("sales",fontsize=8)
_ = axes11[1,1].tick_params(axis='both', which='major', labelsize=6)
_ = axes11[1,1].xaxis.set_ticks(np.arange(0, 365, 100))

# Show fig1.1
plt.show()
plt.close("all")

This shows us the “overall” seasonality pattern across all years: We likely have a strong weekly seasonality pattern that holds across the years, and some monthly seasonality especially towards December.

Monthly & weekly seasonality

Now let’s look at seasonality across days of a month and days of a week. These will likely be the most important seasonality patterns in supermarket sales. There are three ways to look at these plots: First we will group them by year.

Show code

# FIG2: Monthly and weekly seasonality
fig2, axes2 = plt.subplots(2)
_ = fig2.suptitle('Monthly and weekly seasonality, average daily sales in millions')

# Average sales per day of month
_ = sns.lineplot(
  ax = axes2[0],
  x = sales.index.day, 
  y = (sales / 1000000), 
  hue = sales.index.year.astype(str), data=sales)
_ = axes2[0].legend(title = "year", bbox_to_anchor=(1.05, 1.0), fontsize="small", loc='best')
_ = axes2[0].set_xlabel("day of month", fontsize=8)
_ = axes2[0].set_ylabel("sales", fontsize=8)
_ = axes2[0].xaxis.set_ticks(np.arange(1, 32, 6))
_ = axes2[0].xaxis.set_ticks(np.arange(1, 32, 1), minor=True)
_ = axes2[0].yaxis.set_ticks(np.arange(0, 1.25, 0.25))
_ = axes2[0].grid(which='minor', alpha=0.5)
_ = axes2[0].grid(which='major', alpha=1)

# Average sales per day of week
_ = sns.lineplot(
  ax = axes2[1],
  x = (sales.index.dayofweek+1).astype(str), 
  y = (sales / 1000000), 
  hue = sales.index.year.astype(str), data=sales, legend=False)
_ = axes2[1].set_xlabel("day of week", fontsize=8)
_ = axes2[1].set_ylabel("sales", fontsize=8)
_ = axes2[1].yaxis.set_ticks(np.arange(0, 1.25, 0.25))

# Show fig2
plt.show()
plt.close("all")

• The weekly seasonality pattern across days of a week is clear. Sales are lowest in Thursdays, increase and peak at Sundays, then drop on Mondays. The pattern holds in all years.

• The monthly seasonality across days of a month isn’t as strong, but looks considerable. Sales are generally highest at the start of a month, likely because most salaries are paid at the end of a month, though the competition information also says salaries are paid biweekly, in the middle and at the end of each month.

We can look at the same plots grouped by month. The confidence bands are suppressed as they make the plots hard to read.

Show code

# FIG2.1: Monthly and weekly seasonality, colored by month
fig21, axes21 = plt.subplots(2)
_ = fig21.suptitle('Monthly and weekly seasonality, average daily sales in millions')

# Average sales per day of month, colored by month
_ = sns.lineplot(
  ax = axes21[0],
  x = sales.index.day, 
  y = (sales / 1000000), 
  hue = sales.index.month.astype(str), data=sales, errorbar=None)
_ = axes21[0].legend(title = "month", bbox_to_anchor=(1.05, 1.0), fontsize="x-small", loc='best')
_ = axes21[0].set_xlabel("day of month", fontsize=8)
_ = axes21[0].set_ylabel("sales", fontsize=8)
_ = axes21[0].xaxis.set_ticks(np.arange(1, 32, 6))
_ = axes21[0].xaxis.set_ticks(np.arange(1, 32, 1), minor=True)
_ = axes21[0].yaxis.set_ticks(np.arange(0, 1.25, 0.25))
_ = axes21[0].grid(which='minor', alpha=0.5)
_ = axes21[0].grid(which='major', alpha=1)

# Average sales per day of week, colored by month
_ = sns.lineplot(
  ax = axes21[1],
  x = (sales.index.dayofweek+1).astype(str), 
  y = (sales / 1000000), 
  hue = sales.index.month.astype(str), data=sales, errorbar=None, legend=None)
_ = axes21[1].set_xlabel("day of week", fontsize=8)
_ = axes21[1].set_ylabel("sales", fontsize=8)
_ = axes21[1].yaxis.set_ticks(np.arange(0, 1.25, 0.25))

# Show fig2.1
plt.show()
plt.close("all")

This plot shows us the monthly and weekly seasonality pattern generally holds across all months, but December is a notable exception:

• Sales in December are considerably higher roughly from the 13th, due to Christmas approaching. The Christmas peak seems to happen in the 23th, followed by a decline, and another sharp increase in the 30th.

• The sales by day of week are also higher in December for every day, but the pattern of the weekly seasonality is the same.

• We can also see that sales decline very sharply in January 1st, almost to zero, and make a considerably sharp recovery in the 2nd.

And finally, without any grouping: The averages across all years and months.

Show code

# FIG2.2: Monthly and weekly seasonality, average across years
fig22, axes22 = plt.subplots(2)
_ =  fig22.suptitle('Monthly and weekly seasonality, average daily sales in millions')

# Average sales per day of month
_ =  sns.lineplot(
  ax = axes22[0],
  x = sales.index.day, 
  y = (sales / 1000000), 
  data=sales)
_ =  axes22[0].set_xlabel("day of month", fontsize=8)
_ =  axes22[0].set_ylabel("sales", fontsize=8)
_ =  axes22[0].xaxis.set_ticks(np.arange(1, 32, 6))
_ =  axes22[0].xaxis.set_ticks(np.arange(1, 32, 1), minor=True)
_ =  axes22[0].yaxis.set_ticks(np.arange(0, 1.25, 0.25))
_ =  axes22[0].grid(which='minor', alpha=0.5)
_ =  axes22[0].grid(which='major', alpha=1)

# Average sales per day of week
_ =  sns.lineplot(
  ax = axes22[1],
  x = (sales.index.dayofweek+1).astype(str), 
  y = (sales / 1000000), 
  data=sales)
_ =  axes22[1].set_xlabel("day of week", fontsize=8)
_ =  axes22[1].set_ylabel("sales", fontsize=8)
_ =  axes22[1].yaxis.set_ticks(np.arange(0, 1.25, 0.25))

# Show fig22
plt.show()
plt.close("all")

This plot allows us to see the overall monthly seasonality pattern more clearly: Sales are higher at the start of a month, slightly decline until the mid-month payday, slightly increase afterwards, and start peaking again towards the end of a month.

Autocorrelation & partial autocorrelation

Autocorrelation is the correlation of a variable with its own past values. Partial autocorrelation can be thought of as the marginal contribution of each lagged value to autocorrelation, since the lags are likely to hold common information. The patterns in the autocorrelation and partial autocorrelation plots can give us insight into the seasonal patterns.

Show code

# FIG3: ACF and PACF plots
from statsmodels.graphics.tsaplots import plot_acf, plot_pacf
fig3, axes3 = plt.subplots(2)
_ = fig3.suptitle('Autocorrelation and partial autocorrelation, daily sales, up to 54 days')
_ = plot_acf(sales, lags=range(0,55), ax=axes3[0])
_ = plot_pacf(sales, lags=range(0,55), ax=axes3[1], method="ywm")

# Show fig3
plt.show()
plt.close("all")

The sinusoidal pattern in the ACF plot is typical for strong weekly seasonality:

• Sales at T=0 are highly correlated with the sales of the previous day at T-1.

• The correlation declines until T-6, which is the previous value of the next weekday from T=0, and spikes again at T-7, which is the previous value of the same weekday.

• The pattern repeats weekly after T-7, with declining strength.

The PACF plot shows the marginal contribution of each lag to the relationship with the present day value. The partial autocorrelation is highest for lag 1, and significant correlations die out roughly after 14 days. In step 2 of this analysis, we will revisit this plot to derive features from sales lags, after performing time decomposition with model 1.

April 2016 Earthquake

One more time effect we will look at is the April 2016 earthquake, a one-off occurrence we need to adjust for in our model. The plots below show the seasonality of sales in April and May, in every year. The black dashed vertical line marks April 16.

Show code

# FIG6: Zoom in on earthquake: 16 April 2016
april_sales = sales.loc[sales.index.month == 4]
may_sales = sales.loc[sales.index.month == 5]

fig6, axes6 = plt.subplots(2, sharex=True)
_ = fig6.suptitle("Effect of 16 April 2016 earthquake on sales")

# April
_ = sns.lineplot(
  ax = axes6[0],
  x = april_sales.index.day,
  y = april_sales,
  hue = april_sales.index.year.astype(str),
  data = april_sales
)
_ = axes6[0].set_title("April")
_ = axes6[0].legend(title = "year", bbox_to_anchor=(1.05, 1.0), fontsize="small", loc='best')
_ = axes6[0].set_xlabel("days of month")
_ = axes6[0].set_xticks(np.arange(1, 32, 6))
_ = axes6[0].set_xticks(np.arange(1, 32, 1), minor=True)
_ = axes6[0].grid(which='minor', alpha=0.5)
_ = axes6[0].grid(which='major', alpha=1)
_ = axes6[0].axvline(x = 16, color = "black", linestyle = "dashed")

# May
_ = sns.lineplot(
  ax = axes6[1],
  x = may_sales.index.day,
  y = may_sales,
  hue = may_sales.index.year.astype(str),
  data = may_sales, legend=False
)
_ = axes6[1].set_title("May")
_ = axes6[1].set_xlabel("days of month")
_ = axes6[1].set_xticks(np.arange(1, 32, 6))
_ = axes6[1].set_xticks(np.arange(1, 32, 1), minor=True)
_ = axes6[1].grid(which='minor', alpha=0.5)
_ = axes6[1].grid(which='major', alpha=1)

# Show FIG6
plt.show()
plt.close("all")

With the earthquake, the sales sharply increased from the expected seasonal pattern and trend, peaked in April 18, then declined and returned to the seasonal expectations around April 22. It doesn’t seem like sales were higher than normal for the rest of April 2016, or in May 2016. We’ll engineer our earthquake feature accordingly.

Feature engineering: Time & calendar features

Our main candidate for the time decomposition model is a simple linear regression, so we will focus our feature engineering towards that.

• With linear regression, we can model and extrapolate trends and seasonality patterns, as well as adjust for one-off calendar effects.

• Methods that can’t extrapolate outside the training values are not suitable for modeling time effects, even if they are more “advanced”.

Calendar effects & weekly seasonality features

We’ll flag the dates that influence sales considerably with binary or ordinal features. These columns will take non-zero values if a certain calendar effect is present in a date, and zero values otherwise.

Sales decline very sharply every year in January 1st, almost to zero. The sales in January 2nd then recover sharply, which is a huge relative increase for one day, so we will flag both dates with dummy features.

Show code

# New year's day features
df_train["ny1"] = ((df_train.index.day == 1) & (df_train.index.month == 1)).astype(int)


# Set holiday dummies to 0 if NY dummies are 1
df_train.loc[df_train["ny1"] == 1, ["local_holiday", "regional_holiday", "national_holiday"]] = 0

df_train["ny2"] = ((df_train.index.day == 2) & (df_train.index.month == 1)).astype(int)

df_train.loc[df_train["ny2"] == 1, ["local_holiday", "regional_holiday", "national_holiday"]] = 0

We have considerably higher sales in December due to Christmas and the New Years’ Eve.

• For NY’s eve effects, we can simply flag December 30 and 31st with dummy features.

• For Christmas effects, we will create two integer columns that reflect the “strength” of the Christmas effect on sales, based on day of month. The first column will reach its maximum value in December 23rd, and the second will decline as the date moves away from that. For dates outside December 13-27, these features will take a value of zero. This may not be the most sophisticated way to reflect the Christmas effects on our model, but it should be simple and effective.

Show code

# NY's eve features
df_train["ny_eve31"] = ((df_train.index.day == 31) & (df_train.index.month == 12)).astype(int)

df_train["ny_eve30"] = ((df_train.index.day == 30) & (df_train.index.month == 12)).astype(int)

df_train.loc[(df_train["ny_eve31"] == 1) | (df_train["ny_eve30"] == 1), ["local_holiday", "regional_holiday", "national_holiday"]] = 0


# Proximity to Christmas sales peak
df_train["xmas_before"] = 0

df_train.loc[
  (df_train.index.day.isin(range(13,24))) & (df_train.index.month == 12), "xmas_before"] = df_train.loc[
  (df_train.index.day.isin(range(13,24))) & (df_train.index.month == 12)].index.day - 12

df_train["xmas_after"] = 0
df_train.loc[
  (df_train.index.day.isin(range(24,28))) & (df_train.index.month == 12), "xmas_after"] = abs(df_train.loc[
  (df_train.index.day.isin(range(24,28))) & (df_train.index.month == 12)].index.day - 27)

df_train.loc[(df_train["xmas_before"] != 0) | (df_train["xmas_after"] != 0), ["local_holiday", "regional_holiday", "national_holiday"]] = 0

To account for the effect of the April 2016 earthquake, we create a feature similar to the ones for Christmas. The value peaks at April 18th and takes a value of zero for dates outside April 17-22, 2016.

Show code

# Strength of earthquake effect on sales
# April 18 > 17 > 19 > 20 > 21 > 22
df_train["quake_after"] = 0
df_train.loc[df_train.index == "2016-04-18", "quake_after"] = 6
df_train.loc[df_train.index == "2016-04-17", "quake_after"] = 5
df_train.loc[df_train.index == "2016-04-19", "quake_after"] = 4
df_train.loc[df_train.index == "2016-04-20", "quake_after"] = 3
df_train.loc[df_train.index == "2016-04-21", "quake_after"] = 2
df_train.loc[df_train.index == "2016-04-22", "quake_after"] = 1

We previously created binary features to indicate local-regional-national holidays, and special events. There are only a few different events in the dataset, and they differ in nature, so we will break up the events column and create separate binary features for each event. We already created a feature for the earthquake, so we’ll skip that one.

Show code

# Split events, delete events column
df_train["dia_madre"] = ((df_train["event"] == 1) & (df_train.index.month == 5) & (df_train.index.day.isin([8,10,11,12,14]))).astype(int)

df_train["futbol"] = ((df_train["event"] == 1) & (df_train.index.isin(pd.date_range(start = "2014-06-12", end = "2014-07-13")))).astype(int)

df_train["black_friday"] = ((df_train["event"] == 1) & (df_train.index.isin(["2014-11-28", "2015-11-27", "2016-11-25"]))).astype(int)

df_train["cyber_monday"] = ((df_train["event"] == 1) & (df_train.index.isin(["2014-12-01", "2015-11-30", "2016-11-28"]))).astype(int)

df_train = df_train.drop("event", axis=1)

Holidays and events may lead to an increase in sales in advance, so we will create one lag column for each holiday column, and the Mother’s Day event. These features could be tailored more carefully according to each holiday and event, but we’ll keep it simple.

Show code

# Holiday-event leads
df_train["local_lead1"] = df_train["local_holiday"].shift(-1).fillna(0)
df_train["regional_lead1"] = df_train["regional_holiday"].shift(-1).fillna(0)
df_train["national_lead1"] = df_train["national_holiday"].shift(-1).fillna(0)
df_train["diamadre_lead1"] = df_train["dia_madre"].shift(-1).fillna(0)

To capture the weekly seasonality in a simple manner, we’ll create 6 dummy features for days of the week. We won’t create one for Monday, since a 0 value for the other 6 columns means Monday.

Show code

# Days of week dummies
df_train["tuesday"] = (df_train.index.dayofweek == 1).astype(int)
df_train["wednesday"] = (df_train.index.dayofweek == 2).astype(int)
df_train["thursday"] = (df_train.index.dayofweek == 3).astype(int)
df_train["friday"] = (df_train.index.dayofweek == 4).astype(int)
df_train["saturday"] = (df_train.index.dayofweek == 5).astype(int)
df_train["sunday"] = (df_train.index.dayofweek == 6).astype(int)

Now we will aggregate the time features by mean. For local and regional holidays, this will give us fractional values between 0 and 1, which is likely a decent way to reflect local and regional holidays’ effects on national sales.

Show code

# Aggregate time features by mean
time_covars = df_train.drop(
  columns=['id', 'store_nbr', 'family', 'sales', 'onpromotion', 'transactions', 'oil', 'city', 'state', 'store_type', 'store_cluster'], axis=1).groupby("date").mean(numeric_only=True)

Trend & monthly seasonality features

In the exploratory analysis, we saw the overall trend can be modeled linearly, but the rate of increase (the slope of the trend line) declines from the start of 2015. To capture this, we will model a piecewise linear trend with one knot at 01-01-2015.

• The slope of the linear trend will change and decline once, at the knot date. In effect, the result will be two linear trend lines added together.

• It’s possible to add more “turns” to a linear trend with more knots, but one is enough in our case. We want the trend to be simple, and robust against seasonal or cyclical fluctuations.

Show code

# Add piecewise linear trend dummies
time_covars["trend"] = range(1, 1685) # Linear dummy 1

# Knot to be put at period 729
print(time_covars.loc[time_covars.index=="2015-01-01"]["trend"]) 

date
2015-01-01    729
Name: trend, dtype: int64

Show code

# Add second linear trend dummy
time_covars["trend_knot"] = 0
time_covars.iloc[728:,-1] = range(0, 956)

# Check start and end of knot
print(time_covars.loc[time_covars["trend"]>=729][["trend", "trend_knot"]]) 

            trend  trend_knot
date                         
2015-01-01    729           0
2015-01-02    730           1
2015-01-03    731           2
2015-01-04    732           3
2015-01-05    733           4
...           ...         ...
2017-08-11   1680         951
2017-08-12   1681         952
2017-08-13   1682         953
2017-08-14   1683         954
2017-08-15   1684         955

[956 rows x 2 columns]

For the monthly seasonality, we will create Fourier features. This will capture the slight increases and decreases in sales throughout a month, mostly due to proximity to paydays.

• Named after the French mathematician, Fourier series can be used to model any periodic, repeating fluctuation / waveform as sums of numerous sine-cosine pairs.

• We’ll create 5 Fourier pairs (10 columns in total) for 28-period seasonality.

  • Too few pairs may not capture the fluctuations well, while too many run the risk of overfitting.

• December was an exception to the monthly seasonality pattern, but our Christmas and NY features will adjust for that.

Show code

from statsmodels.tsa.deterministic import DeterministicProcess

# Add Fourier features for monthly seasonality
dp = DeterministicProcess(
  index = time_covars.index,
  constant = False,
  order = 0, # No trend feature
  seasonal = False, # No seasonal dummy features
  period = 28, # 28-period seasonality (28 days, 1 month)
  fourier = 5, # 5 Fourier pairs
  drop = True # Drop perfectly collinear terms
)
time_covars = time_covars.merge(dp.in_sample(), how="left", on="date")

# View Fourier features
print(time_covars.iloc[0:5, -10:])

            sin(1,28)  cos(1,28)  sin(2,28)  cos(2,28)  sin(3,28)  cos(3,28)  \
date                                                                           
2013-01-01     0.0000     1.0000     0.0000     1.0000     0.0000     1.0000   
2013-01-02     0.2225     0.9749     0.4339     0.9010     0.6235     0.7818   
2013-01-03     0.4339     0.9010     0.7818     0.6235     0.9749     0.2225   
2013-01-04     0.6235     0.7818     0.9749     0.2225     0.9010    -0.4339   
2013-01-05     0.7818     0.6235     0.9749    -0.2225     0.4339    -0.9010   

            sin(4,28)  cos(4,28)  sin(5,28)  cos(5,28)  
date                                                    
2013-01-01     0.0000     1.0000     0.0000     1.0000  
2013-01-02     0.7818     0.6235     0.9010     0.4339  
2013-01-03     0.9749    -0.2225     0.7818    -0.6235  
2013-01-04     0.4339    -0.9010    -0.2225    -0.9749  
2013-01-05    -0.4339    -0.9010    -0.9749    -0.2225  

Model 1: Time effects decomposition

We will now build our time decomposition linear regression model in Darts, and compare it with some baseline models, as well as some models that are quicker to implement.

Preprocessing

Show code

# Create Darts time series with time features
ts_timecovars = TimeSeries.from_dataframe(
  time_covars, freq = "D", fill_missing_dates = False)

Our target and covariate Series had 1684 rows each, but the Darts TimeSeries we create from them have 1688 dates. This is likely because we have gaps in our original series. We can check this easily in Darts.

Show code

# Scan for gaps
print(ts_timecovars.gaps())

   gap_start    gap_end  gap_size
0 2013-12-25 2013-12-25         1
1 2014-12-25 2014-12-25         1
2 2015-12-25 2015-12-25         1
3 2016-12-25 2016-12-25         1

It seems our data is missing values for December 25th in every year (the data for 2017 ends in August). Darts automatically filled in the missing dates to 1688, but we need to fill the missing values in our target and covariate series.

Show code

# Fill gaps by interpolating missing values
from darts.dataprocessing.transformers import MissingValuesFiller
na_filler = MissingValuesFiller()
ts_sales = na_filler.transform(ts_sales)
ts_timecovars = na_filler.transform(ts_timecovars)

# Scan for gaps again
print(ts_timecovars.gaps())

Empty DataFrame
Columns: [gap_start, gap_end]
Index: []

Our exploratory analysis showed that the seasonal fluctuations in sales become larger over time. We should use a multiplicative decomposition instead of additive.

• In additive decomposition, we decompose a time series as Total = Trend + Seasonality + Remainder.

• In multiplicative decomposition, we decompose a time series as Total = Trend * Seasonality * Remainder.

One way of performing multiplicative decomposition is taking the logarithm of the time series and performing additive decomposition, because if

$Y = T * S * R$,

then

$log(Y) = log(T) + log(S) + log(R)$ .

Show code

# Define functions to perform log transformation and reverse it. +1 to avoid zeroes
def trafo_log(x):
  return x.map(lambda x: np.log(x+1))

def trafo_exp(x):
  return x.map(lambda x: np.exp(x)-1)

We’ll train our time decomposition models on the natural logarithm of the daily sales data from 2013-2016, and validate their prediction performance on 2017. We use an entire year of data for validation to see if our features can capture long-term seasonality and fixed calendar effects.

Show code

# Train-validation split: Pre 2017 vs 2017
y_train1, y_val1 = trafo_log(ts_sales[:-227]), trafo_log(ts_sales[-227:])

Model specification

We’ll compare the performance of our linear regression model against a few other models.

Naive drift and naive seasonal are two baseline models, meant to represent the performance of a very simple prediction.

• Naive drift fits a straight line from the start to the end of the training data, and extrapolates it to make predictions.

• Naive seasonal simply repeats the last K values in the training set.

We’ll also test two seasonal models which require little input. These models cannot use other covariates in their predictions, such as our calendar features.

• Fast Fourier Transform is an algorithm that converts a signal from the time domain to the frequency domain, used in many fields such as engineering and music.

  • In our case, we’ll try to extrapolate the seasonality waves in our data with FFT. The model will first perform detrending with an exponential trend, to try and account for the slope change in 2015.

• Exponential smoothing uses exponentially weighted moving averages to make predictions. The EMA gives more weight to recent observations, and less to older observations.

  • We’ll use the implementation with separate additive trend, seasonal and error components, as well as with trend dampening to try and account for the slope change in 2015. This implementation is also known as the Holt-Winters’ method.
  • The StatsForecastETS implementation we’ll use derives the best parameters automatically from the training data.

And finally, we’ll use our tailored linear regression model with piecewise linear trend dummies, Fourier features for monthly seasonality and dummy features for weekly seasonality & calendar effects. We’ll see how much better it performs compared to models that can be specified and applied quickly.

• All of the covariates we’ll use in this step are known into the future: For example, when predicting the day after our training data, we’ll know what day of the week it will be, or what value our Fourier pairs will take. Darts takes in such covariates as future covariates, and requires their values as inputs for predictions into the future.

• Covariates that are known in the past, but won’t be known in the future are called past covariates. Some Darts models still use these as inputs, but models trained with past covariates have some limitations in future predictions:

  • They can predict up to output_chunk_length time steps without needing values of “future” past covariates.

  • For forecast horizons longer than output_chunk_length, they require “future” values of past covariates. These can be forecasts of the past covariates, of course using forecasts to forecast may lead to compounded errors.

  • We’ll use past covariates in the second step of this analysis.

  • We’ll set output_chunk_length to 15, meaning the model will predict 15 time steps in one go, similar to what the competition requires.

    • The model can still predict forecast horizons longer than 15 time steps, by using the predicted values of the target in auto-regressive fashion, as long as the future and past covariates are supplied. Of course, forecast errors will compound with longer horizons.

Show code

# Import models
from darts.models.forecasting.baselines import NaiveDrift, NaiveSeasonal
from darts.models.forecasting.fft import FFT
from darts.models.forecasting.sf_ets import StatsForecastETS as ETS
from darts.models.forecasting.linear_regression_model import LinearRegressionModel

# Specify baseline models
model_drift = NaiveDrift()
model_seasonal = NaiveSeasonal(K=7) # Repeat the last week of the training data

# Specify FFT model
model_fft = FFT(
  required_matches = {"day", "day_of_week"}, # Try to match the days of month and days of week of the training sequence with the predictions sequence  
  nr_freqs_to_keep = 10, # Keep 10 frequencies
  trend = "exp" # Exponential detrending
)

# Specify ETS model
model_ets = ETS(
  season_length = 7, # Weekly seasonality
  model = "AAA", # Additive trend, seasonality and remainder component
  damped = None # Try both a dampened and non-dampened trend
)

# Specify linear regression model
model_linear1 = LinearRegressionModel(
  lags_future_covariates = [0], # Don't create any future covariate lags
  output_chunk_length = 15 # Predict 15 time steps in one go
  )

Model validation: Predicting 2017 sales

We’ll train our models on 2013-2016 sales data, predict the sales for 2017, and score the predictions with a custom function. We’ll use four regression performance metrics:

• Mean absolute error (MAE),

• Root mean squared error (RMSE),

• Root mean squared log error (RMSLE),

• Mean absolute percentage error (MAPE).

Show code

# Fit models on train data (pre-2017), predict validation data (2017)
_ = model_drift.fit(y_train1)
pred_drift = model_drift.predict(n = 227)

_ = model_seasonal.fit(y_train1)
pred_seasonal = model_seasonal.predict(n = 227)

_ = model_fft.fit(y_train1)
pred_fft = model_fft.predict(n = 227)

_ = model_ets.fit(y_train1)
pred_ets = model_ets.predict(n = 227)

_ = model_linear1.fit(y_train1, future_covariates = ts_timecovars)
pred_linear1 = model_linear1.predict(n = 227, future_covariates = ts_timecovars)

Show code

# Define model scoring function
from darts.metrics import mape, rmse, rmsle, mae
def perf_scores(val, pred, model="drift"):
  
  scores_dict = {
    "MAE": mae(trafo_exp(val), trafo_exp(pred)),
    "RMSE": rmse(trafo_exp(val), trafo_exp(pred)), 
    "RMSLE": rmse(val, pred), 
    "MAPE": mape(trafo_exp(val), trafo_exp(pred))
      }
      
  print("Model: " + model)
  
  for key in scores_dict:
    print(
      key + ": " + 
      str(round(scores_dict[key], 4))
       )
  print("--------")  

# Score models' performance
perf_scores(y_val1, pred_drift, model="Naive drift")
perf_scores(y_val1, pred_seasonal, model="Naive seasonal")
perf_scores(y_val1, pred_fft, model="FFT")
perf_scores(y_val1, pred_ets, model="Exponential smoothing")
perf_scores(y_val1, pred_linear1, model="Linear regression")

Model: Naive drift
MAE: 798635.4611
RMSE: 902091.7355
RMSLE: 0.87
MAPE: 162.8795
--------
Model: Naive seasonal
MAE: 153335.7317
RMSE: 184939.3724
RMSLE: 0.3857
MAPE: 64.8081
--------
Model: FFT
MAE: 184690.6545
RMSE: 228241.9286
RMSLE: 0.3993
MAPE: 60.4031
--------
Model: Exponential smoothing
MAE: 177944.8239
RMSE: 203758.4308
RMSLE: 0.398
MAPE: 76.5003
--------
Model: Linear regression
MAE: 64128.0321
RMSE: 81697.341
RMSLE: 0.1171
MAPE: 9.7622
--------

We see our linear regression model performs much better than the other methods tested.

It’s also notable that the naive seasonal model beats the FFT model in all metrics except MAPE, while ETS scores close to naive seasonal on MAE, RMSE and RMSLE, but much worse on MAPE.

• This is likely because MAPE is a measure of relative error, while the others are measures of absolute error.
  • For example, an absolute error of 2 translates to 2% MAPE if the true value is 100, but it translates to 0.2% MAPE if the true value is 1000.
  • In both cases, the absolute error is the same, but we may argue an absolute error of 2 is more “costly” for the former case.
• In either case, RMSLE is likely the most important metric among these three. Log errors penalize underpredictions more strongly than overpredictions, which could be good for assessing sales predictions, as unfulfilled demand is likely more costly than overstocking. The competition is also scored on RMSLE.
• The forecast horizon plays a huge role in the prediction scores of our models:

  • When the same models are validated on only the last 15 days of our training data, they perform very similarly to one another, and the ETS model has the best performance, even slightly better than our linear regression model with features.

    • The difference in performance is mainly because the models except linear regression can’t take in covariates to capture calendar effects or long-term seasonality. These may not matter much for a short forecast horizon like 15, but with a long forecast horizon like 227, they become more important. The linear regression is able to perform fairly similarly with a forecast horizon of 15 or 227, while the other models do much worse with a longer horizon.

Let’s plot the predictions of our models and compare them visually with the actual values.

Show code

# FIG7: Plot models' predictions against actual values
fig7, axes7 = plt.subplots(4, sharex=True, sharey=True)
_ = fig7.suptitle("Actual vs. predicted sales, time decomposition models, blue = predicted")

# Naive seasonal
_ = trafo_exp(y_val1).plot(ax = axes7[0], label="Actual")
_ = trafo_exp(pred_seasonal).plot(ax = axes7[0], label="Predicted")
_ = axes7[0].set_title("Naive seasonal")
#_ = axes7[0].set_label(fontsize="small")

# FFT
_ = trafo_exp(y_val1).plot(ax = axes7[1], label="Actual")
_ = trafo_exp(pred_fft).plot(ax = axes7[1], label="Predicted")
_ = axes7[1].set_title("FFT")
#_ = axes7[1].set_label(fontsize="small")

# ETS
_ = trafo_exp(y_val1).plot(ax = axes7[2], label="Actual")
_ = trafo_exp(pred_ets).plot(ax = axes7[2], label="Predicted")
_ = axes7[2].set_title("ETS")
#_ = axes7[2].set_label(fontsize="small")

# Linear regression
_ = trafo_exp(y_val1).plot(ax = axes7[3], label="Actual")
_ = trafo_exp(pred_linear1).plot(ax = axes7[3], label="Predicted")
_ = axes7[3].legend("", frameon = False)
_ = axes7[3].set_title("Linear regression")
#_ = axes7[3].set_label(fontsize="small")

# Show FIG7
plt.show()
plt.close("all")

The FFT and ETS models actually did a good job of capturing the weekly seasonality pattern in the data, as the shape and timing of the waves in the predictions are similar to the actual values.

• However, the FFT and ETS models generally overestimated the trend of 2017 sales.

  • This is likely because their methods for modeling the trend is not fully suitable for the piecewise trend in our data.

• The naive seasonal model performed close to FFT and ETS by coincidence, because the 7 days of sales it repeated were not far from the trend of 2017.

• The naive, FFT and ETS models also couldn’t account for the sharp drop in sales in January 1st, as they don’t take in covariates.

• In contrast, the linear model’s trend and seasonality are both on point, and the January 1st drop is adjusted for nicely. The model is not able to match some spikes and troughs fully, which are possibly cyclical in nature. That’s where model 2 will come in.

  • The piecewise linear trend method allows us to adjust for major trend turns in the training data, while keeping the trend line robust for extrapolation into the future.

Rolling crossvalidation

We can further evaluate the performance of our model with rolling crossvalidation: This is the practice of training a model with past data up to a certain time point, predicting the next time point(s), adding them to the training data and repeating the process with increasingly wider training windows until the end of the data.

• We’ll start by training the model on 2013 only, as it makes sense to feed a full year of the seasonality patterns to the model.

• Then we’ll predict the next day, add it to the training data, and repeat the process until the end of the data. This could take a long time for more sophisticated algorithms, but it should be fine for linear regression.

• Keep in mind the performance won’t be fully comparable to our 2017 validation predictions, because those were made with an output chunk length of n=15, while the historical forecasts are computed with an output chunk length of n=1. We do this because the number of timesteps in our data is not divisible by 15, so we can’t obtain a forecast for all timesteps with n=15.

• We’ll also retrieve the residuals for each prediction to perform residual diagnosis.

Show code

# Retrieve historical forecasts and decomposed residuals for 2014-2017
pred_hist1 = model_linear1.historical_forecasts(
  trafo_log(ts_sales), 
  future_covariates = ts_timecovars, 
  start = 365, 
  stride = 1,
  forecast_horizon = 1,
  verbose = False)
res_linear1 = trafo_log(ts_sales[365:]) - pred_hist1

# Score historical forecasts
perf_scores(trafo_log(ts_sales[365:]), pred_hist1, model="Linear regression, historical")

Model: Linear regression, historical
MAE: 89260.1694
RMSE: 114684.7746
RMSLE: 0.3414
MAPE: 16.4487
--------

Show code

# Plot historical forecasts for linear regression
_ = ts_sales.plot(label="Actual")
_ = trafo_exp(pred_hist1).plot(label="Predicted")
_ = plt.title("Time decomposition linear model, historical forecasts")
_ = plt.ylabel("sales")
plt.show()
plt.close("all")

Again, the model captures the trend & seasonality patterns nicely, but the errors due to cyclical effects are even more apparent in 2014 and 2015. If we had modeled the trend non-linearly, the trend term would likely respond strongly to these fluctuations, potentially causing overfitting issues for predictions after the training time period.

Residuals diagnosis & stationarity

We’ll now analyze the residuals from our historical forecasts.

Show code

# Diagnose linear model 1's innovation residuals
from darts.utils.statistics import plot_residuals_analysis, plot_pacf
_ = plot_residuals_analysis(res_linear1)
plt.show()
plt.close("all")

# PACF plot of decomped sales residuals
_ =  plot_pacf(res_linear1, max_lag=56)
_ = plt.title("Partial autocorrelation plot, residuals of linear model 1")
_ = plt.xlabel("Lags")
_ = plt.ylabel("PACF")
_ = plt.xticks(np.arange(0, 56, 10))
_ = plt.xticks(np.arange(0, 56, 1), minor=True)
_ = plt.grid(which='minor', alpha=0.5)
_ = plt.grid(which='major', alpha=1)
plt.show()
plt.close("all")

Ideally, we’d want very small residuals that are normally distributed around zero, representing the truly unpredictable remainder of our models. This is not the case overall, so there’s room for improvement in our predictions.

• Residuals for 2014-2015 especially deviate from zero. This is likely due to the cyclical behavior we observed in this period, though the smaller training data likely plays a role as well.

• The overall distribution of the residuals is not too far from a normal distribution around zero, but there is a huge right tail, indicating some values are strongly underpredicted. These are mostly from 2014-2015, with one particularly huge residual at the start of 2015.

• The PACF plot shows strong partial autocorrelation with lag 1, and a weak one with lag 2, which means we’ll likely include lagged sales features in the second model. There are also small, barely significant partial autocorrelations after lag 2.

A stationary time series is one with values independent of the observation time, free of trend and seasonality. Ideally, we’d expect our decomposed residuals to be a stationary, white noise series constantly fluctuating around zero. A related concept is the presence or lack of a unit root.

• A deterministic trend is a linear trend plus small fluctuations around it with constant variance. It is fairly stable over time. Removing the trend yields a stationary error term with constant variance. Such a series is trend-stationary.

• A unit root, i.e. a stochastic trend is a linear trend plus a random walk process: The series fluctuates considerably around the linear trend, and can shift from it permanently. Removing the linear trend won’t make the series stationary, but differencing will (more on this later). Such a series is difference-stationary.

We can test stationarity and the presence of a unit root in the residuals with two statistical tests. First is the Kwiatkowski-Phillips-Schmidt-Shin test.

Show code

# Import KPSS and ADF tests
from darts.utils.statistics import stationarity_test_kpss, stationarity_test_adf

print(
  "KPSS test p-value: " + 
  stationarity_test_kpss(res_linear1)[1].astype(str)
) # Null rejected = data has unit root

KPSS test p-value: 0.01

C:\Users\PC\Documents\Work\DataScience\GitHub\KaggleStoreSales\venv\lib\site-packages\statsmodels\tsa\stattools.py:2018: InterpolationWarning:

The test statistic is outside of the range of p-values available in the
look-up table. The actual p-value is smaller than the p-value returned.

The null hypothesis for this version of the KPSS test is stationarity around a constant value, while the alternative suggests the presence of a unit root. With a p-value smaller than 0.01, the null is rejected, so the test suggests our residuals have a unit root.

Second is the Augmented Dickey-Fuller test.

Show code

print(
  "ADF test p-value: " +
  stationarity_test_adf(res_linear1)[1].astype(str)
) # Null rejected = data has no unit root

ADF test p-value: 2.492249770117031e-05

The null hypothesis for the ADF test is the presence of a unit root, while the alternative hypothesis is lack of one. With a very small p-value, the null is rejected and the alternative is accepted, so the test suggests our residuals do not have a unit root, conflicting with the KPSS test.

• This conflict is likely due to the unequal variance in the residuals, especially for values between 2014-2015.

Time decomposition

We will train our second model on the time-decomposed sales, which are the residuals of the first model.

• Since 2013-2016 will serve as our training data, we need to fit the first model on 2013-2016, and retrieve the fitted residuals to serve as the time-decomposed training data for the second model, but Darts doesn’t allow predictions on the training data (which makes sense, as this is normally a mistake if the goal is to evaluate model performance).

• Instead, it’s possible to use the model_linear1.residuals() method, which returns the residuals from rolling crossvalidation as performed previously, but this will lead to higher residuals for the earlier days in our data, as they will be retrieved from models fitted on small portions of our available data, without seeing all the examples of the time patterns we want to remove.

• Since we are ultimately interested in improving our model’s performance in 2017 predictions, and we already validated model 1 on 2017, we’ll perform the final time decomposition in sklearn, which allows predictions on the training data.

Show code

# Perform time decomposition in Sklearn

# Retrieve Darts target and features with filled gaps
sales = ts_sales.pd_series()
time_covars = ts_timecovars.pd_dataframe()

# Train-test split
y_train, y_val = trafo_log(sales[:-227]), trafo_log(sales[-227:])
x_train, x_val = time_covars.iloc[:-227,], time_covars.iloc[-227:,]

# Fit & predict on 13-16, retrieve residuals
from sklearn.linear_model import LinearRegression
model_decomp = LinearRegression()
model_decomp.fit(x_train, y_train)
pred1 = model_decomp.predict(x_train)
res1 = y_train - pred1

# Predict on 17, retrieve residuals
pred2 = model_decomp.predict(x_val)
res2 = y_val - pred2

# Concatenate residuals to get decomposed sales
sales_decomp = pd.concat([res1, res2])

# Concatenate predictions to get model 1 predictions
preds_model1 = pd.Series(np.concatenate((pred1, pred2)), index = sales_decomp.index)

Now, preds_model1 are the fitted values of model 1 on 2013-2016, and its predictions for 2017. sales_decomp are the fitted residuals for 2013-2016 and the prediction residuals for 2017. We’ll fit model 2 on sales_decomp, and add preds_model1 to its predictions to achieve our final hybrid predictions.

Exploratory analysis 2: Lags & covariates

Now we will explore the sales lags & covariate time series as possible predictors for our second model.

Show code

# Aggregate daily covariates from df_train: oil, onpromotion, transactions
covars = df_train.groupby("date").agg(
 { "oil": "mean",
  "onpromotion": "sum",
  "transactions": "mean"}
  )

# Merge decomposed sales and covariates
sales_covariates = covars.merge(sales_decomp.rename("sales"), on = "date", how = "left")

Covariates stationarity & differencing

Oil, onpromotion and transactions are time series that we will consider as predictors of sales.

• However, we performed time decomposition on sales. Our covariate series may display their own time effects and seasonalities. We can use stationarity tests to quickly check for this.

• If a covariate is non-stationary, a quick way to possibly make it stationary is differencing: Subtracting each value from the value of the previous period. This will give us the change in each period, instead of the absolute value.

Show code

# Test stationarity of covariates
from statsmodels.tsa.stattools import kpss, adfuller

print(
  "KPSS test p-value: " + 
  kpss(sales_covariates["oil"])[1].astype(str)
) # Null rejected = data is non-stationary

print(
  "ADF test p-value: " +
  adfuller(sales_covariates["oil"])[1].astype(str)
) # Null accepted = data has unit root

KPSS test p-value: 0.01
ADF test p-value: 0.8037342071858674

Both the KPSS and ADF test suggest the oil time series is non-stationary, so we’ll difference it. If one covariate is differenced, it makes sense to difference the others as well.

Show code

# Difference the covariates
from sktime.transformations.series.difference import Differencer
diff = Differencer(lags = 1)
sales_covariates[
  ['oil', 'onpromotion', 'transactions']] = diff.fit_transform(
    sales_covariates[['oil', 'onpromotion', 'transactions']])
print(sales_covariates)

               oil  onpromotion  transactions   sales
date                                                 
2013-01-01  0.0000       0.0000        0.0000 -0.7649
2013-01-02  0.0000       0.0000    92445.0000  0.0696
2013-01-03 -0.1507       0.0000   -14711.0000  0.2949
2013-01-04  0.1330       0.0000      -10.0000  0.1230
2013-01-05  0.0236       0.0000    15079.0000  0.1126
...            ...          ...           ...     ...
2017-08-11  0.2176    6262.0000     6487.0000 -0.1177
2017-08-12 -0.3277   -5867.0000      376.0000 -0.3223
2017-08-13 -0.3277     971.0000    -3934.0000 -0.2947
2017-08-14 -0.3277   -1240.0000     -545.0000 -0.1223
2017-08-15 -0.0161    2562.0000     1113.0000 -0.0757

[1684 rows x 4 columns]

Now our covariate columns all represent the change in each covariate compared to the previous period.

Sales features

Let’s evaluate features derived from sales lags as possible predictors. We’ll start with autocorrelation plots.

Show code

from statsmodels.graphics.tsaplots import plot_pacf as pacf_tsa
from sktime.utils.plotting import plot_lags
from scipy.stats import pearsonr, spearmanr

Show code

# FIG8: Sales ACF - PACF plot
fig8, axes8 = plt.subplots(2, sharex=True)
_ = fig8.suptitle('Autocorrelation and partial autocorrelation, decomposed sales, up to 14 days')

_ = plot_acf(
  sales_covariates["sales"], lags=np.arange(0, 15, 1, dtype=int), ax=axes8[0], marker=".")
_ = pacf_tsa(
  sales_covariates["sales"], lags=np.arange(0, 15, 1, dtype=int), ax=axes8[1], method="ywm", marker=".")

_ = axes8[0].xaxis.set_ticks(np.arange(0, 15, 1, dtype=int), minor=True)
_ = axes8[0].xaxis.set_ticks(np.arange(0, 15, 7, dtype=int))
_ = axes8[0].grid(which='minor', alpha=0.5)
_ = axes8[1].xaxis.set_ticks(np.arange(0, 15, 1, dtype=int), minor=True)
_ = axes8[1].xaxis.set_ticks(np.arange(0, 15, 7, dtype=int))
_ = axes8[1].grid(which='minor', alpha=0.5)

# Show fig8
plt.show()
plt.close("all")

As we saw in model 1’s residuals analysis, sales lag 1 makes a strong contribution to the partial autocorrelation, while lags 2 to 5 make weak but significant contributions.

Linear correlation may not always do justice to the relationship between two variables, so let’s look at the scatterplots of sales with its first 9 lags.

Show code

# FIG9: Sales lag scatterplots
fig9, axes9 = plot_lags(
  sales_covariates["sales"], lags = [1,2,3,4,5,6,7,8,9],
  suptitle = "Sales lags")
  
# Show fig9
plt.show()
plt.close("all")

It doesn’t look like any lag after 1 displays a considerably different relationship with sales.

We can also consider a rolling feature, such as a moving average.

• Since we know lag 1 is the most important lag, an exponentially weighted moving average makes sense. This type of moving average puts more weight on recent lags, and less on older lags.

• A 5-day period makes sense, as the significance of partial autocorrelations die out after lag 5.

Show code

# Calculate 5-day exponential moving average of sales lags
sales_covariates["sales_ema5"] = sales_covariates["sales"].rolling(
  window = 5, 
  min_periods = 1, 
  center = False, # T0 is the end of the window instead of the center
  closed = "left", # Exclude sales at T0 from the calculation
  win_type = "exponential").mean()

Show code

# Plot sales_ema5 vs sales  
_ = sns.regplot(
  data = sales_covariates,
  x = "sales_ema5",
  y = "sales"
)
_ = plt.title("Relationship of sales and 5-day exponential moving average of sales")
plt.show()
plt.close("all")

We can see sales EMA5 is a good linear predictor of sales, similar to lag 1. Let’s compare lag 1’s correlation with sales to EMA5’s correlation with sales.

Show code

# Correlation of sales and lag 1
pearsonr(sales_covariates["sales"], sales_covariates["sales"].shift(1).fillna(method="bfill")) 

PearsonRResult(statistic=0.7905238105754431, pvalue=0.0)

Show code

# Correlation of sales and ema5
pearsonr(sales_covariates["sales"], sales_covariates["sales_ema5"]) 

PearsonRResult(statistic=0.7582545348904933, pvalue=8.454075723e-315)

Lag 1 comes out as the slightly stronger linear predictor, but EMA5 is a good one too. We can include both as features.

Show code

# Keep sales EMA5 as a feature, but calculate it with T0 included, as Darts will ask for at least 1 past covariate lag
sales_covariates["sales_ema5"] = sales_covariates["sales"].rolling(
  window = 5, 
  min_periods = 1, 
  center = False, # T0 is the end of the window instead of the center
  closed = "right", # Include T0 in the calculation
  win_type = "exponential").mean()

Oil features

The oil price change of one day is not likely to mean much by itself, and the effect of oil prices on the economy likely takes some time to manifest.

• Therefore, we’ll just consider moving averages of the oil price as features.

• These won’t be exponentially weighted averages, as recent oil prices likely require time to make their effect on sales. It’s even possible oil prices with a certain degree of lag are the more important predictors.

Show code

# Calculate oil MAs
oil_ma = sales_covariates.assign(
  oil_ma7 = lambda x: x["oil"].rolling(window = 7, center = False, closed = "left").mean(),
  oil_ma14 = lambda x: x["oil"].rolling(window = 14, center = False, closed = "left").mean(),
  oil_ma28 = lambda x: x["oil"].rolling(window = 28, center = False, closed = "left").mean(),
  oil_ma84 = lambda x: x["oil"].rolling(window = 84, center = False, closed = "left").mean(),
  oil_ma168 = lambda x: x["oil"].rolling(window = 168, center = False, closed = "left").mean(),
  oil_ma336 = lambda x: x["oil"].rolling(window = 336, center = False, closed = "left").mean(),
)

Show code

# FIG10: Regplots of oil moving averages & sales
fig10, axes10 = plt.subplots(3,2, sharey=True)
_ = fig10.suptitle("Oil price change moving averages & decomposed sales")

# MA7
_ = sns.regplot(
  ax = axes10[0,0],
  data = oil_ma,
  x = "oil_ma7",
  y = "sales"
)
_ = axes10[0,0].set_xlabel("weekly MA")
_ = axes10[0,0].annotate(
    'Corr={:.2f}'.format(
      spearmanr(oil_ma["oil_ma7"], oil_ma["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# MA14
_ = sns.regplot(
  ax = axes10[0,1],
  data = oil_ma,
  x = "oil_ma14",
  y = "sales"
)
_ = axes10[0,1].set_xlabel("biweekly MA")
_ = axes10[0,1].annotate(
    'Corr={:.2f}'.format(
      spearmanr(oil_ma["oil_ma14"], oil_ma["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# MA28
_ = sns.regplot(
  ax = axes10[1,0],
  data = oil_ma,
  x = "oil_ma28",
  y = "sales"
)
_ = axes10[1,0].set_xlabel("monthly MA")
_ = axes10[1,0].annotate(
    'Corr={:.2f}'.format(
      spearmanr(oil_ma["oil_ma28"], oil_ma["sales"], nan_policy='omit')[0]
      ),
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# MA84
_ = sns.regplot(
  ax = axes10[1,1],
  data = oil_ma,
  x = "oil_ma84",
  y = "sales"
)
_ = axes10[1,1].set_xlabel("quarterly MA")
_ = axes10[1,1].annotate(
    'Corr={:.2f}'.format(
      spearmanr(oil_ma["oil_ma84"], oil_ma["sales"], nan_policy='omit')[0]
      ),
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# MA168
_ = sns.regplot(
  ax = axes10[2,0],
  data = oil_ma,
  x = "oil_ma168",
  y = "sales"
)
_ = axes10[2,0].set_xlabel("semi-annual MA")
_ = axes10[2,0].annotate(
    'Corr={:.2f}'.format(
      spearmanr(oil_ma["oil_ma168"], oil_ma["sales"], nan_policy='omit')[0]
      ),
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# MA336
_ = sns.regplot(
  ax = axes10[2,1],
  data = oil_ma,
  x = "oil_ma336",
  y = "sales"
)
_ = axes10[2,1].set_xlabel("annual MA")
_ = axes10[2,1].annotate(
    'Corr={:.2f}'.format(
      spearmanr(oil_ma["oil_ma336"], oil_ma["sales"], nan_policy='omit')[0]
      ),
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# Show fig10
plt.show()
plt.close("all")

We don’t see particularly strong relationships between oil price change MAs and sales, though the monthly MA is the most notable one with a correlation of -0.19, so we will use that as a feature.

Since we need 28 past periods to calculate this MA for a 29th period, we’ll have missing values for the first 28 periods in our data. After some iteration, these were interpolated decently with a second-order spline interpolation.

Show code

# Keep monthly oil MAs as a feature
sales_covariates["oil_ma28"] = sales_covariates["oil"].rolling(window = 28, center = False).mean()

# Spline interpolation for missing values
sales_covariates["oil_ma28"] = sales_covariates["oil_ma28"].interpolate(
  method = "spline", order = 2, limit_direction = "both")

# Check quality of interpolation
_ = sales_covariates["oil"].plot()
_ = sales_covariates["oil_ma28"].plot()
_ = plt.title("Oil price change and its 28-day moving average, first 28 MAs interpolated")
plt.show()
plt.close("all")

Onpromotion features

Onpromotion is the number of items that are on sale, aggregated across all stores and categories. We’ll start exploring the relationship between onpromotion and sales using a cross-correlation plot: The correlations of sales at time 0, with lags and leads of onpromotion. Since the supermarket chain controls the number of items that will go on sale, it may be feasible to use leading values of onpromotion as predictors.

Show code

# Cross-correlation of sales & onpromotion
_ = plt.xcorr(
  sales_covariates["sales"], sales_covariates["onpromotion"], usevlines=True, maxlags=56, normed=True)
_ = plt.grid(True)
_ = plt.ylim([-0.1, 0.1])
_ = plt.xlabel("onpromotion lags / leads")
_ = plt.title("Cross-correlation, decomposed sales & onpromotion change")
plt.show()
plt.close("all")

Single lags / leads of onpromotion don’t display any significant correlations with sales, though there is a repeating pattern of likely six days. Let’s look at the scatterplots of sales at time 0, with four of the more relatively significant lags (lag 0 being the present value of onpromotion).

Show code

# FIG12: Onpromotion lags 0 1 6 7
fig12, axes12 = plt.subplots(2,2, sharey=True)
_ = fig12.suptitle("Onpromotion change lags & decomposed sales")

# Lag 0
_ = sns.regplot(
  ax = axes12[0,0],
  x = sales_covariates["onpromotion"],
  y = sales_covariates["sales"]
)
_ = axes12[0,0].set_xlabel("lag 0")
_ = axes12[0,0].annotate(
    'Corr={:.2f}'.format(
      spearmanr(sales_covariates["onpromotion"], sales_covariates["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )
      
# Lag 1
_ = sns.regplot(
  ax = axes12[0,1],
  x = sales_covariates["onpromotion"].shift(1),
  y = sales_covariates["sales"]
)
_ = axes12[0,1].set_xlabel("lag 1")
_ = axes12[0,1].annotate(
    'Corr={:.2f}'.format(
      spearmanr(sales_covariates["onpromotion"].shift(1), sales_covariates["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# Lag 6
_ = sns.regplot(
  ax = axes12[1,0],
  x = sales_covariates["onpromotion"].shift(6),
  y = sales_covariates["sales"]
)
_ = axes12[1,0].set_xlabel("lag 6")
_ = axes12[1,0].annotate(
    'Corr={:.2f}'.format(
      spearmanr(sales_covariates["onpromotion"].shift(6), sales_covariates["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# Lag 7
_ = sns.regplot(
  ax = axes12[1,1],
  x = sales_covariates["onpromotion"].shift(7),
  y = sales_covariates["sales"]
)
_ = axes12[1,1].set_xlabel("lag 7")
_ = axes12[1,1].annotate(
    'Corr={:.2f}'.format(
      spearmanr(sales_covariates["onpromotion"].shift(7), sales_covariates["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# Show fig12
plt.show()
plt.close("all")

Again, no significant relationship between single onpromotion lags and sales. Let’s consider moving averages.

Show code

# FIG11: Onpromotion MAs
fig11, axes11 = plt.subplots(2,2)
_ = fig11.suptitle("Onpromotion change moving averages & decomposed sales")

# Calculate MAs without min_periods = 1
onp_ma = sales_covariates.assign(
  onp_ma7 = lambda x: x["onpromotion"].rolling(window = 7, center = False, closed = "left").mean(),
  onp_ma14 = lambda x: x["onpromotion"].rolling(window = 14, center = False, closed = "left").mean(),
  onp_ma28 = lambda x: x["onpromotion"].rolling(window = 28, center = False, closed = "left").mean(),
  onp_ma84 = lambda x: x["onpromotion"].rolling(window = 84, center = False, closed = "left").mean()
)

# MA7
_ = sns.regplot(
  ax = axes11[0,0],
  data = onp_ma,
  x = "onp_ma7",
  y = "sales"
)
_ = axes11[0,0].set_xlabel("weekly MA")
_ = axes11[0,0].annotate(
    'Corr={:.2f}'.format(
      spearmanr(onp_ma["onp_ma7"], onp_ma["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# MA14
_ = sns.regplot(
  ax = axes11[0,1],
  data = onp_ma,
  x = "onp_ma14",
  y = "sales"
)
_ = axes11[0,1].set_xlabel("biweekly MA")
_ = axes11[0,1].annotate(
    'Corr={:.2f}'.format(
      spearmanr(onp_ma["onp_ma14"], onp_ma["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# MA28
_ = sns.regplot(
  ax = axes11[1,0],
  data = onp_ma,
  x = "onp_ma28",
  y = "sales"
)
_ = axes11[1,0].set_xlabel("monthly MA")
_ = axes11[1,0].annotate(
    'Corr={:.2f}'.format(
      spearmanr(onp_ma["onp_ma28"], onp_ma["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# MA84
_ = sns.regplot(
  ax = axes11[1,1],
  data = onp_ma,
  x = "onp_ma84",
  y = "sales"
)
_ = axes11[1,1].set_xlabel("quarterly MA")
_ = axes11[1,1].annotate(
    'Corr={:.2f}'.format(
      spearmanr(onp_ma["onp_ma84"], onp_ma["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# Show fig11
plt.show()
plt.close("all")

The monthly and quarterly MA of onpromotion seems to be slightly positively correlated with sales. We’ll use the monthly MA as a feature, as we would have to fill in too many missing values for the quarterly MA.

Show code

# Keep monthly onpromotion MAs as a feature
sales_covariates["onp_ma28"] = sales_covariates["onpromotion"].rolling(window = 28, center = False).mean()

# Spline interpolation for missing values
sales_covariates["onp_ma28"] = sales_covariates["onp_ma28"].interpolate(
  method = "spline", order = 2, limit_direction = "both")

# Check quality of interpolation
_ = sales_covariates["onpromotion"].plot()
_ = sales_covariates["onp_ma28"].plot()
_ = plt.title("Onpromotion change and its 28-day MA, first 28 interpolated")
plt.show()
plt.close("all")

The daily changes in onpromotion are practically zero roughly until Q1 2014. Afterwards, changes in onpromotion became increasingly larger, which probably hints at a macroeconomic shift in 2014. The cyclicality in our data becomes especially prominent in 2014 as well.

Transactions features

Finally, we have the number of transactions, aggregated across all stores.

• This is most likely to be a current predictor of sales, i.e. days with higher transactions are likely to be days with higher sales. We wouldn’t know the number of transactions in advance, so this may not be useful. Still, recent levels of transactions may help us predict future sales.

• The transactions data is only available for the training dataset (up to mid-August 2017), and not for the testing dataset (15 days after the training dataset), which has implications for its use in part 2 of this analysis. Lag 15 is marked with a red dashed vertical line in the cross-correlation plot below.

Show code

# Cross-correlation of sales & transactions change
_ = plt.xcorr(
  sales_covariates["sales"], sales_covariates["transactions"], usevlines=True, maxlags=56, normed=True)
_ = plt.grid(True)
_ = plt.ylim([-0.15, 0.15])
_ = plt.axvline(x = -14, color = "red", linestyle = "dashed")
_ = plt.xlabel("transactions lags / leads")
_ = plt.title("Cross-correlation, decomposed sales & transactions change")
plt.show()
plt.close("all")

Again, single transactions lags don’t display strong correlations with sales. Unlike onpromotion, there doesn’t seem to be a repeating pattern either. Lags 0-3 are relatively stronger, so let’s see their scatterplots.

Show code

# FIG14: Transactions lags 0-3
fig14, axes14 = plt.subplots(2,2, sharey=True)
_ = fig14.suptitle("Transactions change lags & decomposed sales")

# Lag 0
_ = sns.regplot(
  ax = axes14[0,0],
  x = sales_covariates["transactions"],
  y = sales_covariates["sales"]
)
_ = axes14[0,0].set_xlabel("lag 0")
_ = axes14[0,0].annotate(
    'Corr={:.2f}'.format(
      spearmanr(sales_covariates["transactions"], sales_covariates["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )
      
# Lag 1
_ = sns.regplot(
  ax = axes14[0,1],
  x = sales_covariates["transactions"].shift(1),
  y = sales_covariates["sales"]
)
_ = axes14[0,1].set_xlabel("lag 1")
_ = axes14[0,1].annotate(
    'Corr={:.2f}'.format(
      spearmanr(sales_covariates["transactions"].shift(1), sales_covariates["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# Lag 2
_ = sns.regplot(
  ax = axes14[1,0],
  x = sales_covariates["transactions"].shift(2),
  y = sales_covariates["sales"]
)
_ = axes14[1,0].set_xlabel("lag 2")
_ = axes14[1,0].annotate(
    'Corr={:.2f}'.format(
      spearmanr(sales_covariates["transactions"].shift(2), sales_covariates["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# Lag 3
_ = sns.regplot(
  ax = axes14[1,1],
  x = sales_covariates["transactions"].shift(3),
  y = sales_covariates["sales"]
)
_ = axes14[1,1].set_xlabel("lag 3")
_ = axes14[1,1].annotate(
    'Corr={:.2f}'.format(
      spearmanr(sales_covariates["onpromotion"].shift(3), sales_covariates["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# Show fig14
plt.show()
plt.close("all")

As we expected, the present value of transactions have a weak, but relatively stronger positive relationship with the value of sales compared to the lags.

• However, the correlation and the line is impacted by a few outliers with particularly low or high values for transactions, while the relationship seems even weaker for the vast majority of observations.

• This likely tells us that a huge drop / increase in transactions is likely to affect sales a bit, but usual fluctuations in transactions doesn’t tell much about the sales.

Let’s consider moving averages of changes in transactions, as a possible indicator of recent sales activity.

Show code

# FIG13: Transactions MAs
fig13, axes13 = plt.subplots(2,2)
_ = fig13.suptitle("Transactions change moving averages & decomposed sales")

# Calculate MAs without min_periods = 1
trns_ma = sales_covariates.assign(
  trns_ma7 = lambda x: x["transactions"].rolling(window = 7, center = False, closed = "left").mean(),
  trns_ma14 = lambda x: x["transactions"].rolling(window = 14, center = False, closed = "left").mean(),
  trns_ma28 = lambda x: x["transactions"].rolling(window = 28, center = False, closed = "left").mean(),
  trns_ma84 = lambda x: x["transactions"].rolling(window = 84, center = False, closed = "left").mean()
)

# MA7
_ = sns.regplot(
  ax = axes13[0,0],
  data = trns_ma,
  x = "trns_ma7",
  y = "sales"
)
_ = axes13[0,0].set_xlabel("weekly MA")
_ = axes13[0,0].annotate(
    'Corr={:.2f}'.format(
      spearmanr(trns_ma["trns_ma7"], trns_ma["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# MA14
_ = sns.regplot(
  ax = axes13[0,1],
  data = trns_ma,
  x = "trns_ma14",
  y = "sales"
)
_ = axes13[0,1].set_xlabel("biweekly MA")
_ = axes13[0,1].annotate(
    'Corr={:.2f}'.format(
      spearmanr(trns_ma["trns_ma14"], trns_ma["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# MA28
_ = sns.regplot(
  ax = axes13[1,0],
  data = trns_ma,
  x = "trns_ma28",
  y = "sales"
)
_ = axes13[1,0].set_xlabel("monthly MA")
_ = axes13[1,0].annotate(
    'Corr={:.2f}'.format(
      spearmanr(trns_ma["trns_ma28"], trns_ma["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# MA84
_ = sns.regplot(
  ax = axes13[1,1],
  data = trns_ma,
  x = "trns_ma84",
  y = "sales"
)
_ = axes13[1,1].set_xlabel("quarterly MA")
_ = axes13[1,1].annotate(
    'Corr={:.2f}'.format(
      spearmanr(trns_ma["trns_ma84"], trns_ma["sales"], nan_policy='omit')[0]
      ), 
      xy=(.6, .9), xycoords="axes fraction", bbox=dict(alpha=0.5)
      )

# Show fig13
plt.show()
plt.close("all")

The MAs are more significant than single lags, though the correlations are still likely impacted by extreme values. Still, these features could help our model adjust for extreme increases / decreases in sales. We’ll keep the weekly MA as it has the strongest linear relationship, and will create the fewest number of missing values to fill in.

Show code

# Keep weekly transactions MA
sales_covariates["trns_ma7"] = sales_covariates["transactions"].rolling(window = 7, center = False).mean()

# Backwards linear interpolation
sales_covariates["trns_ma7"] = sales_covariates["trns_ma7"].interpolate("linear", limit_direction = "backward")

# Check quality of interpolation
_ = sales_covariates["transactions"].plot()
_ = sales_covariates["trns_ma7"].plot()
_ = plt.title("Transactions change and its 7-day moving average, first 7 MAs interpolated")
plt.show()
plt.close("all")

We can see the sharp drop in transactions on January 1st, and the recovery on January 2nd, reflected both in the time series plot of transactions changes, and its weekly moving average.

Model 2: Lags & covariates

Preprocessing

We repeat some preprocessing steps for model 2.

Show code

# Drop original covariates & target
lags_covars = sales_covariates.drop(
  ['oil', 'onpromotion', 'transactions', 'sales'], axis = 1)

# Make Darts TS with selected covariates
ts_lagscovars = TimeSeries.from_dataframe(
  lags_covars, freq="D")

# Make Darts TS with decomposed sales
ts_decomp = TimeSeries.from_series(sales_decomp.rename("sales"), freq="D")

# Make Darts TS with model 1 predictions
ts_preds1 = TimeSeries.from_series(preds_model1.rename("sales"), freq="D")

# Fill gaps in TS
ts_decomp = na_filler.transform(ts_decomp)
ts_lagscovars = na_filler.transform(ts_lagscovars)

# Train-test split (pre-post 2017)
y_train2, y_val2 = ts_decomp[:-227], trafo_log(ts_sales[-227:])
x_train2, x_val2 = ts_lagscovars[:-227], ts_lagscovars[-227:]

In model 1, our features were trend dummies, Fourier pairs and binary / ordinal features for calendar effects, none of which required scaling. In this model, we have covariates of different natures and value ranges. For example, the oil price changes are often fractional, while changes in transactions are often in the thousands. The former reflects a unit price, the latter reflects a count. Therefore, we’ll scale and center our features to avoid domination by features with large values.

Show code

# Scale covariates
from sklearn.preprocessing import StandardScaler
from darts.dataprocessing.transformers import Scaler
scaler = Scaler(StandardScaler())
x_train2 = scaler.fit_transform(x_train2)
x_val2 = scaler.transform(x_val2)

# Combine back scaled covariate series
ts_lagscovars_scaled = x_train2.append(x_val2)

Model specification

We’ll use the same baseline models as we did in model 1. We’ll compare them with an ARIMA model, another linear regression, and a random forest model.

• ARIMA models combine autoregressive models (AR) and moving average models (MA).

  • An AR model of order p uses the linear combination of the past p values of a series to forecast its future values.

  • A MA model of order q performs regression on the past q forecast errors of a series to forecast its future values. It can be thought of as a weighted moving average of the past q forecast errors.

  • The order d indicates the order of differentiation applied to the target series.

  • This implementation of the ARIMA model also takes in covariates, which will be used as linear predictors just like in linear regression.

  • We’ll use an AutoARIMA process, which will derive and apply the best ARIMA parameters from the data. The best ARIMA model will be chosen by minimizing the Corrected Akaike Information Criterion (AICc), which reflects the tradeoff between variance explained by the model, and model complexity.

• A random forest builds an N number of decision trees, each with an M number of randomly selected predictors out of all the predictors in the dataset. The predictions of each tree are combined with a voting system, yielding an ensemble prediction.

  • Tree-based models can perform feature selection to a degree, capture non-linear relationships, and discover interactions between predictors.

  • They can’t extrapolate a relationship outside the value ranges in the training set, so we didn’t consider them for the time decomposition model. This inability may still impact the second model’s ability to predict future values.

  • We could also include a more sophisticated gradient boosting algorithm such as XGBoost, but these often require extensive hyperparameter tuning to perform at their best, so we will skip it. In contrast, the random forest has fewer hyperparameters, and the default settings often do fine without needing tuning.

  • Darts uses the Scikit-learn implementation of the random forest regression model, RandomForestRegressor.

    • This implementation uses all features to build all N trees by default, which I believe is misleading: Randomly subsetting the features to build each tree is a big part of how RF models combat overfitting. Taking that away just leaves us with an ensemble of trees built with sampled data, without any pruning, likely prone to overfitting.

    • Remember to set max_features which will be randomly chosen to build each tree to avoid this pitfall. Here we use a common heuristic for regression, which is the square root of the number of features, though it may not change much as we have few features.

Show code

# Import models
from darts.models.forecasting.auto_arima import AutoARIMA
from darts.models.forecasting.linear_regression_model import LinearRegressionModel
from darts.models.forecasting.random_forest import RandomForest

# Specify new seasonal baseline model
model_seasonal = NaiveSeasonal(K=5)

# Specify AutoARIMA model
model_arima = AutoARIMA(
  start_p = 1,
  max_p = 5,
  start_q = 0,
  max_q = 5,
  d = 0,
  seasonal = False, # Don't include seasonal terms
  stationary = True, # The data is stationary, don't difference
  information_criterion = 'aicc', # Minimize AICc to choose best model
  trace = True, # Print the tuning iterations
  random_state = 1923
)

# Specify second linear regression model
model_linear2 = LinearRegressionModel(
  lags = [-1], # Use one target lag
  lags_future_covariates = [0], # No future covariate lags
  lags_past_covariates = [-1], # Use lag 1 of past covariates
  output_chunk_length = 15 # Predict 15 steps in one go
  )

# Specify random forest model  
model_forest = RandomForest(
  lags = [-1],
  lags_future_covariates = [0],
  lags_past_covariates = [-1],
  output_chunk_length = 15,
  n_estimators = 500, # Build 500 trees,
  max_features = "sqrt", # Use a subset of features for each tree
  oob_score = True, # Score trees on out-of-bag samples
  random_state = 1923,
  n_jobs = 20
  )

Similar to model 1, our naive seasonal model will repeat the last 5 values in the training data. K=7 was the most significant seasonality period before time decomposition. Now it appears K=5 is the most significant period for autoregression.

• To determine the search spaces for the AR and MA components in the AutoARIMA model, we look at the ACF and PACF plots of decomposed sales (see Sales features).

  • The ACF plot shows exponential / sinusoidal decline, and the last significant spike in the PACF plot is at lag 2, though small but significant partial correlations persist up to lag 5.

  • Accordingly, we’ll consider AR components from 1 to 5, and MA components from 0 to 5.

  • We specify a differencing order of 0, as our target is already decomposed of time effects. We also do not perform detrending, or include seasonal components in our ARIMA models.

• Our linear and random forest regression models will take in sales lag 1, sales EMA5, oil MA28, onpromotion MA28 and transactions MA7 as predictors.

  • Sales and transaction MAs will be past covariates: Their future values won’t be used for predictions. It makes sense to validate our model like this, as we won’t know the future values of these variables for our test data predictions in part 2.

Model validation: Predicting 2017 sales

Let’s train our models on the decomposed sales of 2013-2016, predict the values for 2017, and add these to the 2017 predictions of model 1 to retrieve our final, hybrid model predictions.

Show code

# Specify future and past covariates
fut_cov = ['oil_ma28', 'onp_ma28']
past_cov = ['sales_ema5', 'trns_ma7']

# Fit models on train data (pre-2017), predict validation data (2017)
_ = model_drift.fit(y_train2)
pred_drift2 = model_drift.predict(n = 227) + ts_preds1[-227:]

_ = model_seasonal.fit(y_train2)
pred_seasonal2 = model_seasonal.predict(n = 227) + ts_preds1[-227:]

_ = model_arima.fit(
  y_train2, 
  future_covariates = ts_lagscovars_scaled[fut_cov]
)

pred_arima = model_arima.predict(
  n = 227, 
  future_covariates = ts_lagscovars_scaled[fut_cov]
  ) + ts_preds1[-227:]

_ = model_linear2.fit(
  y_train2, 
  future_covariates = ts_lagscovars_scaled[fut_cov],
  past_covariates = ts_lagscovars_scaled[past_cov]
  )
  
pred_linear2 = model_linear2.predict(
  n = 227, 
  future_covariates = ts_lagscovars_scaled[fut_cov],
  past_covariates = ts_lagscovars_scaled[past_cov]
  ) + ts_preds1[-227:]

_ = model_forest.fit(
  y_train2, 
  future_covariates = ts_lagscovars_scaled[fut_cov],
  past_covariates = ts_lagscovars_scaled[past_cov]
  )

pred_forest = model_forest.predict(
  n = 227, 
  future_covariates = ts_lagscovars_scaled[fut_cov],
  past_covariates = ts_lagscovars_scaled[past_cov]
  ) + ts_preds1[-227:]

Performing stepwise search to minimize aicc

 ARIMA(1,0,0)(0,0,0)[0] intercept   : AICC=-2658.061, Time=0.26 sec
 ARIMA(0,0,0)(0,0,0)[0] intercept   : AICC=-1209.321, Time=0.11 sec

 ARIMA(0,0,1)(0,0,0)[0] intercept   : AICC=-2047.417, Time=0.16 sec
 ARIMA(0,0,0)(0,0,0)[0]             : AICC=-1211.332, Time=0.12 sec

 ARIMA(2,0,0)(0,0,0)[0] intercept   : AICC=-2669.324, Time=0.33 sec

 ARIMA(3,0,0)(0,0,0)[0] intercept   : AICC=-2688.389, Time=0.45 sec

 ARIMA(4,0,0)(0,0,0)[0] intercept   : AICC=-2688.234, Time=0.47 sec

 ARIMA(3,0,1)(0,0,0)[0] intercept   : AICC=-2698.913, Time=0.63 sec

 ARIMA(2,0,1)(0,0,0)[0] intercept   : AICC=-2504.471, Time=0.63 sec

 ARIMA(4,0,1)(0,0,0)[0] intercept   : AICC=-2685.232, Time=0.73 sec

 ARIMA(3,0,2)(0,0,0)[0] intercept   : AICC=-2701.060, Time=0.75 sec

 ARIMA(2,0,2)(0,0,0)[0] intercept   : AICC=-2694.612, Time=0.67 sec

 ARIMA(4,0,2)(0,0,0)[0] intercept   : AICC=-2696.165, Time=0.83 sec

 ARIMA(3,0,3)(0,0,0)[0] intercept   : AICC=-2689.299, Time=0.84 sec

 ARIMA(2,0,3)(0,0,0)[0] intercept   : AICC=-2691.841, Time=0.58 sec

 ARIMA(4,0,3)(0,0,0)[0] intercept   : AICC=-2693.147, Time=0.90 sec

 ARIMA(3,0,2)(0,0,0)[0]             : AICC=-2703.529, Time=0.65 sec

 ARIMA(2,0,2)(0,0,0)[0]             : AICC=-2701.749, Time=0.56 sec

 ARIMA(3,0,1)(0,0,0)[0]             : AICC=-2701.738, Time=0.58 sec

 ARIMA(4,0,2)(0,0,0)[0]             : AICC=-2701.644, Time=0.77 sec

 ARIMA(3,0,3)(0,0,0)[0]             : AICC=-2700.784, Time=0.81 sec

 ARIMA(2,0,1)(0,0,0)[0]             : AICC=-2686.749, Time=0.58 sec

 ARIMA(2,0,3)(0,0,0)[0]             : AICC=-2693.831, Time=0.50 sec

 ARIMA(4,0,1)(0,0,0)[0]             : AICC=-2688.787, Time=0.68 sec

 ARIMA(4,0,3)(0,0,0)[0]             : AICC=-2697.922, Time=0.83 sec

Best model:  ARIMA(3,0,2)(0,0,0)[0]          
Total fit time: 14.413 seconds

Show code

# Score models' performance
perf_scores(y_val2, pred_drift2, model="Naive drift")
perf_scores(y_val2, pred_seasonal2, model="Naive seasonal")
perf_scores(y_val2, pred_arima, model="ARIMA")
perf_scores(y_val2, pred_linear2, model="Linear")
perf_scores(y_val2, pred_forest, model="Random forest")

Model: Naive drift
MAE: 110648.8939
RMSE: 135351.1619
RMSLE: 0.1986
MAPE: 15.2878
--------
Model: Naive seasonal
MAE: 138273.0834
RMSE: 162484.9224
RMSLE: 0.2206
MAPE: 20.7798
--------
Model: ARIMA
MAE: 57659.4119
RMSE: 78381.3524
RMSLE: 0.1078
MAPE: 8.3075
--------
Model: Linear
MAE: 55895.8078
RMSE: 78738.332
RMSLE: 0.1072
MAPE: 7.8976
--------
Model: Random forest
MAE: 59505.884
RMSE: 79361.7179
RMSLE: 0.113
MAPE: 8.9219
--------

Remember that these are hybrid predictions: Each prediction is the combined result of the linear regression time decomposition model, and the respective lags & covariates models.

• As a refresher, model 1’s standalone performance scores on 2017 data:

  --------
  Model: Linear regression
  MAE: 64128.0321
  RMSE: 81697.341
  RMSLE: 0.1171
  MAPE: 9.7622
  --------
  

• Combining model 1 with a baseline model doesn’t increase prediction errors all that much. This is because our data had strong time effects, so getting model 1 right did most of the job.

• Combining model 1 with a second linear regression model yields the best performance in terms of MAE, RMSLE and MAPE, followed by ARIMA and random forest.

  • The improvement over model 1 is not big. This implies the effects of our covariates aren’t too important compared to the trend, seasonality and calendar effects.

• ARIMA yields the best performance on RMSE scores, though the scores are very close.

• RMSE punishes larger errors more severely due to the squaring applied, while RMSLE punishes underpredictions more severely due to the log transformation. So the ARIMA likely makes a little less error for larger values, while also underpredicting compared to linear regression.

Let’s see the predictions plotted against the actual values.

Show code

# FIG15: Plot models' predictions against actual values
fig15, axes15 = plt.subplots(3, sharex=True, sharey=True)
_ = fig15.suptitle("Actual vs. predicted sales, hybrid models, blue = predicted")

# ARIMA
_ = trafo_exp(y_val2).plot(ax = axes15[0], label="Actual")
_ = trafo_exp(pred_arima).plot(ax = axes15[0], label="Predicted")
_ = axes15[0].set_title("Linear regression + ARIMA")

# Linear regression
_ = trafo_exp(y_val2).plot(ax = axes15[1], label="Actual")
_ = trafo_exp(pred_linear2).plot(ax = axes15[1], label="Predicted")
_ = axes15[1].set_title("Linear regression + Linear regression")

# Random forest
_ = trafo_exp(y_val2).plot(ax = axes15[2], label="Actual")
_ = trafo_exp(pred_forest).plot(ax = axes15[2], label="Predicted")
_ = axes15[2].legend("", frameon = False)
_ = axes15[2].set_title("Linear regression + Random forest")

# Show fig15
plt.show()
plt.close("all")

The performance of the 3 hybrids are similar overall, and it’s hard to see big differences from the plots.

• The random forest did a better job of adjusting to some spikes and troughs especially at the start of 2017, possibly because it is able to model non-linear relationships and find interactions between predictors.

• Still, certain large spikes are partially unaccounted for, especially the sharp recovery after January 1st, and the spikes at the start of April and May 2017. These may be due to factors not documented in our dataset. Or they may be related to commemorations for the earthquake in April 2016.

ARIMA model summary

Before carrying on with our best hybrid model, which is linear + linear, we can view the model summary of the ARIMA model for some insight.

The AutoARIMA process has chosen an ARIMA(3, 0, 2) model, which uses the last 3 lags and the last 2 forecast errors as predictors, along with our covariates. Covariates x1 and x2 refer to our future covariates, oil_ma28 and onp_ma28 respectively.

Show code

print(model_arima.model.summary())

                               SARIMAX Results                                
==============================================================================
Dep. Variable:                      y   No. Observations:                 1461
Model:               SARIMAX(3, 0, 2)   Log Likelihood                1359.814
Date:                Fri, 19 May 2023   AIC                          -2703.628
Time:                        17:39:39   BIC                          -2661.333
Sample:                             0   HQIC                         -2687.851
                               - 1461                                         
Covariance Type:                  opg                                         
==============================================================================
                 coef    std err          z      P>|z|      [0.025      0.975]
------------------------------------------------------------------------------
x1            -0.0195      0.009     -2.068      0.039      -0.038      -0.001
x2             0.0103      0.002      5.203      0.000       0.006       0.014
ar.L1          0.7103      0.168      4.237      0.000       0.382       1.039
ar.L2          0.6730      0.268      2.516      0.012       0.149       1.197
ar.L3         -0.4157      0.118     -3.526      0.000      -0.647      -0.185
ma.L1         -0.0052      0.164     -0.032      0.975      -0.326       0.316
ma.L2         -0.6918      0.154     -4.503      0.000      -0.993      -0.391
sigma2         0.0091      0.000     50.172      0.000       0.009       0.009
===================================================================================
Ljung-Box (L1) (Q):                   0.71   Jarque-Bera (JB):              2542.43
Prob(Q):                              0.40   Prob(JB):                         0.00
Heteroskedasticity (H):               0.98   Skew:                             0.79
Prob(H) (two-sided):                  0.83   Kurtosis:                         9.27
===================================================================================

Warnings:
[1] Covariance matrix calculated using the outer product of gradients (complex-step).

The AIC score is not meaningful on its own without being compared to AIC scores of other models, though smaller values (including negative) indicate a better trade-off between variance explained and model complexity. This was the ARIMA model with the lowest AIC score among those tried.

• The first two lags have positive coefficients (ar.L1 and ar.L2), so a higher value for lags 1 & 2 means a higher prediction for sales at T=0.

• Lag 3 and the forecast errors 1 & 2 (ma.L1, ma.L2) have a negative effect on the prediction. If lag 3 or the forecast errors increase, the prediction for T=0 is pulled back.

  • The first forecast error’s effect is not significant, with p = 0.975, and the 95% confidence interval for the coefficient value ranging from roughly -0.3 to 0.3.

• The p-value for the Ljung-Box test is 0.4, so the null hypothesis of residual independence is accepted. The residuals of our model are not correlated, so applying a third autoregressive model on them is not likely to improve predictions.

• The p-value for the heteroskedasticity test is 0.83, so the residuals meet the equal variance assumption. The model makes similar errors across the entire predicted time series, which means its performance is consistent.

• The null hypothesis for the Jarque-Bera test is zero skew & zero excess kurtosis in the distribution of the tested data. With a p-value of 0, the null hypothesis is rejected, and we can see the residuals have a skew of 0.79 and kurtosis of 9.27, compared to 0 and 3 respectively for a normal distribution.

  • A high kurtosis distribution can be thought of as a narrower and taller distribution compared to normal, which likely means our residuals are packed tightly around zero (or possibly a non-zero value).
  • The positive skew means the residuals distribution has a right tail, i.e. is skewed to the left. This likely means the model overpredicts more than it underpredicts, which is preferable.

Backtesting / Historical forecasts

We can retrieve the historical forecasts for model 2 as we did for model 1, and add them to model 1’s historical forecasts to retrieve historical forecasts of the hybrid model.

• Keep in mind we performed time decomposition by fitting model 1 on 2013-2016 and retrieving the fitted residuals for 2013-2016. This means these historical forecasts shouldn’t be interpreted as reliable out-of-sample validation.

• 2017 does not suffer from this issue, as 2017 time decomposed sales were predicted out-of-sample.

• As we did before, we’ll start with the first 365 days, and expand our training window by 1 day in each iteration.

Show code

# Retrieve historical forecasts & residuals for linear + random forest

# Fit scaler on 2013, transform 2013+
scaler.fit(ts_lagscovars[:365])
ts_lagscovars_hist = scaler.transform(ts_lagscovars)


# Predict historical forecasts for 2014-2017
pred_hist2 = model_linear2.historical_forecasts(
  ts_decomp, 
  future_covariates = ts_lagscovars_hist[fut_cov],
  past_covariates = ts_lagscovars_hist[past_cov],
  start = 365, 
  stride = 1,
  forecast_horizon = 1,
  verbose = False) + ts_preds1[365:]

# Retrieve residuals for 2014-2017
res_hist2 = trafo_log(ts_sales[365:]) - pred_hist2

# Score historical forecasts for linear + random_forest
perf_scores(trafo_log(ts_sales[365:]), pred_hist2, model="Linear + linear")

Model: Linear + linear
MAE: 48026.0543
RMSE: 78790.2799
RMSLE: 5.9406
MAPE: 8.4835
--------

Show code

# Plot historical forecasts for random forest
_ = ts_sales.plot(label="Actual")
_ = trafo_exp(pred_hist2).plot(label="Predicted")
_ = plt.title("Linear regression + linear regression hybrid model, historical forecasts")
_ = plt.ylabel("sales")
plt.show()
plt.close("all")

With model 1, the overall historical forecasts performed considerably worse than the 2017 predictions. With the hybrid, the historical forecasts perform fairly close to the 2017 predictions, even a bit better on MAE.

• Again, this is partly misleading due to the way we performed time decomposition, but the improvements from accounting for cyclical effects in 2014-2015 are likely a factor too.

• As we see from the historical forecasts plot, the cyclical fluctuations in 2014 and H1 2015, which were mostly unaccounted for in model 1, are matched nicely with the hybrid model. The spikes and troughs in the rest of the data are also matched better.

• The RMSLE score is absurdly high due to a few values erratically predicted as zero in mid-2014 (I do not know why, possibly due to the very sharp drop in sales that precedes these predictions). These errors are heavily penalized due to the nature of the RMSLE metric.

Residuals diagnosis & stationarity

We can diagnose the residuals of the hybrid model and test their stationarity. The erratic predictions in mid-2014 will yield massive residuals, so we’ll clip them to 0.5 to get interpretable plots.

Show code

# Clip high end of residuals
res_hist2_plot = res_hist2.map(lambda x: np.clip(x, a_min = None, a_max = 0.5))

# Residuals diagnosis
_ = plot_residuals_analysis(res_hist2_plot)
plt.show()
plt.close("all")

# PACF plot
_ = plot_pacf(res_hist2_plot, max_lag=56)
_ = plt.title("Partial autocorrelation plot, hybrid model innovation residuals")
_ = plt.xlabel("Lags")
_ = plt.ylabel("PACF")
_ = plt.xticks(np.arange(0, 56, 10))
_ = plt.xticks(np.arange(0, 56, 1), minor=True)
_ = plt.grid(which='minor', alpha=0.5)
_ = plt.grid(which='major', alpha=1)
plt.show()
plt.close("all")

Overall, the residuals seem much closer to stationary.

• There are a few particularly large residuals (besides the erratic predictions’ residuals clipped to 0.5), especially in 2014-2015.

• The distribution of residuals is more centered around 0, with smaller extreme values. This time, the extreme values are well-balanced between positive and negative. It seems the model overpredicts a bit more than it underpredicts (as the ARIMA summary suggested), which is preferable to the opposite.

• The PACF plot still displays significant autocorrelations, mainly for lags 1-3. It’s possible the linear regression couldn’t account for the exact effect of lagged values, which may be non-linear in nature. Or, the ACF and PACF calculations are strongly impacted by the few erratic predictions in mid-2014.

Show code

# Stationarity tests on hybrid model residuals
print(
  "KPSS test p-value: " + 
  stationarity_test_kpss(res_hist2)[1].astype(str)
) # Null accepted = data is stationary

print(
  "ADF test p-value: " + 
  str(stationarity_test_adf(res_hist2)[1])
) # Null rejected = no unit root

KPSS test p-value: 0.07112705469666736
ADF test p-value: 2.2979600604438747e-08

The stationarity tests are no longer in conflict:

• The p-value for the KPSS test is larger than 0.05 , leading us to accept the null hypothesis of stationarity around a constant, though we are still close to non-stationarity.

• The p-value for the ADF test is practically 0, rejecting the null hypothesis of the presence of an unit root.

Conclusion

In part 1 of this analysis, we were able to predict national daily sales, aggregated across all categories and stores, pretty well.

• In the end, a hybrid model combining a linear regression for the time effects, and a second linear regression with covariates and lagged values for the remainder, gave us the best results. Using an ARIMA or random forest for the second model yielded very similar performance.

• However, the second model’s relative improvement was not big, as the first model already performed very well.

• This means the time & seasonality effects in our data were much stronger than any “unexpected” cyclicality, which is what we’d intuitively expect for supermarket sales.

• Still, there was considerable cyclicality in 2014 and H1 2015, which showed up as relatively large values in the historical residuals of model 1. Applying autoregression & utilizing covariates in model 2 helped us adjust for these.

• Autoregression is more “reactive” than predictive: It won’t predict an initial unexpected fluctuation in advance, but that fluctuation will enter the model as a predictor in the next period, and allow the model to adjust predictions for the following periods.

  • However, this may also lead to erratic crashes in predictions, as we saw in the historical forecasts of model 2.

In part 2, we will attempt to make a competition submission, predicting the daily sales for every store & category combination. That makes 33 categories across 54 stores, a total of 1782 series to generate a prediction for.

• In part 1, with only one series to predict, we were able to perform manual exploratory analysis and carefully tailor our models, especially for time decomposition. It’s also generally easier to predict more aggregated series, as they are less noisy: The national sales likely fluctuate much less compared to sales in one store, one category.

• In part 2, we will have to come up with a strategy that will simplify our work while still yielding good predictions. Some of the 33 product categories likely have very different seasonality patterns. We’ll likely rely on more “automatic” models that are able to derive time effects & seasonality, such as AutoARIMA, or neural networks that can be trained “globally” on multiple time series to try and model them in one go.

• We’ll also explore the topic of hierarchical reconciliation: Making sure our predictions are coherent within the levels of the time series hierarchy. There are several methods to reconcile predictions in a hierarchy, and it can even improve prediction accuracy.

Any comments or suggestions are welcome.