README.md

In gouthaman87/FoodDemandApp: Food demand forecast

Food Demand Forecasting

Project Intro:

The goal of this project is to provide forecast values for the number of orders for meal/center combos to a food delivery service. We have a total of 3548 meal/center combos (i.e. 77 centers & 51 meals), meaning that 3548 time series models will have to be fitted. This technique in business environment is also known as Scalable Forecasting. The project details & materials are available on the following link https://datahack.analyticsvidhya.com/contest/genpact-machine-learning-hackathon-1/.

In this github repo the methodology, R scripts, submission files, time series models & machine learning models used to tackle this problem can be found.

PS: This is not the ONLY method to tackle this problem. However, this is one way to tackle this problem.

The following time series models were used to forecast the number of orders: Naive, Seasonal Naive, Drift, Simple Exponential Smoothing (SES), Holt’s Linear, Additive Damped Trend, STL Decomposition with ETS, ARIMA, Dynamic Harmonic Regression, Time Series Linear Regression, CROSTON, SBA, Neural Network & Ensemble Model. These were not the only models used to forecast; there are plenty of other models that can be used for this problem. However, the computational time to train the models must be considered.

The software used for this work was R and the following R packages were used to forecast: fpp3 : This package contains several time series model functions and other time series related functions. This package includes tsibble, tsibbbledata, fable and feasts packages. tsintermittent : This package is used to find demand patterns and also contains intermittent time series related function. tidyverse : This package is used for data wrangling with pipe operators (tidy work flow).

Step 01: Data Preperation & Finding Demand Patterns.

In this stage, data wrangling steps were performed. This data was then transformed to time series data (i.e. to tsibble object: this is a special type of data that handles time series models). The following shows the data wrangling steps:

The first few weeks of transactions for Center #24 (which can be seen in the plot below) were removed as there were no transactions available for several weeks after these first few weeks. However, there were continuous transactions after this time period, hence these were used to fit the model. The following plot shows aggregated number of orders for Center #24:

A complete meal/center combos time series data was made i.e. meaning that some meal/center combos had some missing weeks; new entries were created for those missing entries by replacing the number of orders with 0 and filled with previous values for other variables (emailer_for_promotion, homepage_featured, base_price, checkout_price). For example, the following time series data shows that after the 4th week there is data missing up to 7th week. Table 2 shows the completed data with the new entries for those missing weeks (i.e. weeks 5 & 6).

Some Meal/Center combos showed less transactions i.e. fewer than 20 transactions. These combos were separated and forecast models were fitted specifically for them using Cross Validation (CV) method.

Finally, the following demand patterns for each meal/center combo, based on the number of orders, were identified: Smooth, Erratic, Lumpy & Intermittent. The method used here to identify demand patterns was Syntetos, Boylan and Croston (2004) (SBC). Identifying these demand patterns meant that different times series models can be applied to them specifically. For example, Croston & SBA were suitable for intermittent demand patterns. Hence, this helped to increase the accuracy of forecasting. R script for Step 01 can be found in 00-GT-Data Prep & Discovery.R & demand pattern table can be found in following path: data\processed\ts_categorization.csv.

Smooth:

The smooth demand pattern shows a regular demand and regular time. i.e. This type of products can be sold every day or every week.

Erratic:

The erratic demand pattern shows regularity in time but the selling quantity varied dramatically. i.e. This type of products can be sold everyday or every week however, for example, one day it may sell 3 in quantity whereas, another day it could sell 100 in quantity.

Intermittent:

The intermittent demand pattern shows irregularity in time and regularity in quantity pattern. i.e. This type of product is sold for the first week then for several weeks would not be sold but at the end, the same amount of product is sold.

Lumpy:

The lumpy demand pattern shows irregularity in time and irregularity in the quantity pattern. This particular type of demand pattern was difficult to forecast no matter what type of time series models was used. A solution for this type of product is to have a safety stock.

The plots below show what each demand pattern looks like:

The above plot showed that in the data, a majority of time series combos fell under Smooth & Erratic. This meant that a regular time series model such as ARIMA, ETS etc. would have suited well. However, advanced models such as Croston & SBA were fitted in order to tackle the intermittent & lumpy demand pattern. The Cross Validation method had been used to fit No Demand (i.e. transactions less than 20) combos.

Step 02: Find suitable models for Smooth & Erratic Patterns.

In this stage, time series models for Smooth & Erratic demand patterns were fitted i.e. First, the data was split into train & test (last 10 weeks) data for each meal/center combo. The following time series models were then fitted for each train data of meal/center combo. Here, number of orders (Y) was log transformed to impose a positive constraint by adding unit 1 to overcome log(0) infinity issue.

1. Naive

The Naive model sets all future values the same as the last observation value.

2. Seasonal Naive

The seasonal naive model was used for seasonal data. In this situation, each future value is equal to the last value from the same season.

3. Drift

The drift model is a variation of Naive model, allowing forecast to increase or decrease over time.

4. Simple Exponential Smooth

The simple exponential smooth model was used when a clear trend or seasonal pattern was not identified.

5. Holt’s Linear

The Holt’s Linear model is an extended version of SES allowing to forecast with trend.

6. Damped Additive Trend

The Damped Additive Trend model is an extended version of Holt’s Linear, allowing trend to change over time with a damped parameter phi.

7. Forecasting with Decomposition

The decomposition model was used to decompose the time series using Seasonal and Trend decomposition using Loess (STL) method. ETS method was then used to forecast the seasonally adjusted data, after which the seasonal component was added to the forecast data.

8. ARIMA

The ARIMA models explain the autocorrelations in the data. ARIMA function was used from fpp3 R package to fit the ARIMA models, which automatically selected the ARMA orders using min AICC. The explanation and theory of ARIMA Models can be found on the following link: https://otexts.com/fpp3/arima.html.

9. Dynamic Harmonic Regression

The Dynamic Harmonic Regression model allows the inclusion of other data, such as base price, checkout price, email promotion & homepage featured. The explanation and theory of Dynamic Harmonic Regression Models can be found on the following link: https://otexts.com/fpp3/dynamic.html.

10. Time Series Regression

The time series regression model is used to forecast the dependent variable Y assuming that it has a linear relationship with other independent variables X. i.e. in this situation, an assumption was made when forecasting that the number of orders and checkout price, base price, emailer for promotion & homepage featured had a linear relationship.

11. Ensemble

The ensemble model simply uses several different models at the same time and calculates average value of the resulting forecasts. For example, here, ARIMA, SES & Decomposition models were used together to calculate the average forecast value.

R script for Step 02 can be found on 01-GT-forecast_model-smooth.R. The forecast values from fitted models can be found in the following path: data\processed\smooth_forecast_tbl.csv

Step 03: Find suitable models for Intermittent & Lumpy Patterns.

In this stage, time series models for Intermittent & Lumpy demand patterns were fitted i.e. Firstly, the data was split into train & test (last 10 weeks) data for each meal/center combo. The following time series models ( CROSTON, SBA, SES, ARIMA & Ensemble ) were then fitted.

1. CROSTON

The Croston model is the most suitable method for slow moving products (intermittent). The theory behind this method can be found on the following link: https://cran.r-project.org/web/packages/tsintermittent/index.html.

2. SBA

The SBA model is another variant / improved version of Croston method. Theory and research papers for this method can be found on the following link: https://cran.r-project.org/web/packages/tsintermittent/index.html.

3. Ensemble for Intermittent

Here, CROSTON, SBA, ARIMA & SES models were used together to calculate the average forecast value.

R script for Step 03 can be found in 02-GT-intermittent forecast.R. The forecast values from fitted models can be found in data\processed\inter_forecast_tbl.csv.

Step 04: Find Suitable Forecast Models.

In this stage, suitable models for each meal/center combo were then found within the fitted models. i.e. The accuracy for each meal/center/model combo (i.e. 3548x7 time series models ) was calculated in order to select a suitable model for each meal/center combo. The accuracy metrics used here was RMSLE. RMSLE was used as it was the metrics used for this competition. However, personally RMSSE was preferred as this was used in the M5 forecasting competition. The following plots show the accuracy (RMSLE) density plot, giving an overview of how many meal/center combos achieved high or low accuracy:

The above plot shows that high accuracy (i.e. less than 1 RMSLE) was achieved for a majority of meal/center combos. However, there were a few meal/center combos which revealed low accuracy; for these low accuracy combos, other advanced time series/machine learning models could be fitted to increase the forecast accuracy. Furthermore, the majority of combos with low accuracy were Lumpy/Intermittent which meant that for these type of combos, rather than fitting advance models, the focus should be on safety stock calculations.

The following plot shows the number of meal/center combos against the models, chosen by minimum RMSLE.

The above plot shows that ARIMA, Dynamic Harmonic Regression & Time Series Regression models are the most fitted models for meal/center combos. R script for Step 04 can be found in 03-Accuracy calculation.R. Suitable model lists selected using minimum RMSLE can be found in the following path: data\processed\suitable_model_tbl.csv.

Step 05: Final Step.

In this stage, forecast for each meal/center combo was performed with the new data provided by the hackathon competition (data\raw\test\test_QoiMO9B.csv). The time series models used here can be found in Step 04 (data\processed\suitable_model_tbl.csv). The forecast final output submission file can be found in output\GT_Submission_date. The following plot shows how the forecast values look in the train/test split stage and how they will look in the future by using the most suitable model. For example, the plot below shows how the forecast values look in the test period (Week 136 - 145) and then in the future period (146 - 155) in Center 10 and Meal 1885 , using the most suitable model Decomposition Model.

R script for Step 05 can be found in 05-GT-Pipeline.R. Forecast - related useful links: https://otexts.com/fpp3/ : The forecasting book. https://kourentzes.com/forecasting/2014/06/23/intermittent-demand-forecasting-package-for-r/ : Intermittent forecasting related theory. https://frepple.com/blog/demand-classification/ : Demand patterns explanations. https://cran.r-project.org/web/packages/tsintermittent/index.html : Intermittent forecasting related theory.

gouthaman87/FoodDemandApp documentation built on Jan. 3, 2023, 12:09 p.m.