M5ForecastingCompetition-Submission

Top 4% in final results for accuracy, 9% for Uncertainty. Submission for M5 Forecasting Competition of Walmart Data.
https://www.kaggle.com/c/m5-forecasting-accuracy
https://www.kaggle.com/c/m5-forecasting-uncertainty/

Data provided by Walmart of the daily sales and price of over 30 000 products over 5 years. Two different sets of predictions were required:

Accuracy competition - point forecasts - scored using weighted mean squared scaled error
Uncertainty competition - interval forecasts - creating the 1%.5%.33%,50%,67%,95%,99% intervals for every prediction, scored using scaled pinball loss for each quantile

Files:

submission_joshli_m5acccuracy.ipynb is the full submission notebook, originally made on kaggle.com using the data here

Uncertainty Stream contains all the files needed to produce the uncertainty interval listed above (including the point forecasts downloaded from submission_joshli_m5acccuracy.ipynb)

From this competition I learned:

• Uncertainty intervals are as equally important as good point predictions, extremely useful to organizations avoiding stockouts
• Significantly less tools were available to make good quantile predictions, and only 909 submissions were made compared to 5558 for point forecasts (accuracy competition) My submission was ranking 74th (top 9%) with room significant for improvement.

From the discussions of top submissions I decided to investigate:

Innovation State Space Model (based on Exponential Smoothing)
Quantile Gradient Boosting regression
Seasonal ARIMA models

So far quantile gradient boosting regression has been the most promising on some backtests and a new project involving real business data from a local boutique.