NOTE THAT THIS APPROACH DID NOT YIELD ANY GOOD SCORES:D!
In the M5 competitions I was in a team called YJ kagglers and responsible for forecasting with statistical models.
In the previous M4 competition, weighted ensemble of statistical models won the 3rd place. It sounded to me that 1) a time-series forecasting is still very hard for even a sophisticated machine learning model, and 2) simple statistical models may work better than GBDT or NN models.
My approach was therefore an ensemble of statistical forecasting models enpowered by sktime, prophet, and statsmodels.
The following models are used:
- Prophet
- ARIMA model
- Theta method model
- Exponential Smoothing model
- Naive model with 'seasonal_last' strategy
- Naive model with 'last' strategy
These models took the univariate (demand) input and forecast the subsequent 28 days demands on a dept_id and store_id basis, and then divided the predicted sum into the prediction of each item based on the original ratio to the sum. Only for items with high-variance, models were fitted individually.
- Running models on individual IDs (> 30,000). It just took very long to complete, but might be better given the diversity of time-series of items' demands.
- Uncertainty competition: tried it but confidence intervals yielded by statistical models weren't great in terms of the score.
This code (run_accuracy.py) uses the paralell processing, so running it locally may fail.
To reproduce the results, using docker is recommended.
How to execute with docker is the following. Note that docker / docker-compose are already in your local environment.
cp project.env .env # step1: copy environment variables to .env
nano .env # step2: edit, if necessary
In .env...
INPUT_DIR: input directory OUTPUT_DIR: output directory
Executing a script is done within the docker container.
docker-compose up -d --build # step3: build the image to initiate container
docker exec -it katsu-m5-conda bash
(base) root@0f8f3ab4a6a1:/analysis# cd script
(base) root@0f8f3ab4a6a1:/analysis/script# python run_accuracy.py 'SktimeEnsemble' 'sub'
Note that the M5 data need to be in the 'input' folder.