Learnings from Kaggle's Forecasting Competitions - Githubissues

kiccho1101 / paper

Interesting Papers, Articles, and Github repositories.

6 stars 0 forks source link

Learnings from Kaggle's Forecasting Competitions #7

Open kiccho1101 opened 4 years ago

kiccho1101 commented 4 years ago

What's this paper about?

Introduces 6 time series competitions held by Kaggle.
Background: Real-life business forecasting tasks on Kaggle platform has been largely ignored by the academic community.

What they did

Identify 6 Kaggle forecasting competitions.
Analyze the competition datasets in terms of time series characteristics.
Benchmark Kaggle solutions.
Review top-scoring solutions.

6 Forecasting competitions

Walmart Store Sales Forecasting
Rossmann Store Sales
Walmart Sales in Stormy Weather
Wikipedia Web Traffic Time Series Forecasting
Corporación Favorita Grocery Sales Forecasting
Recruit Restaurant Visitor Forecasting

kiccho1101 commented 4 years ago

Competitions details

Screen Shot 2020-05-04 at 12 43 39

kiccho1101 commented 4 years ago

Analyzing Datasets

Time series summarization

Time series data can be summarized by 6 features.
f = F(F1, F2, F3, F4, F5, F6).
F1: Spectral entropy. (forecastability.)
F2: Strength of trend.
F3: Strength of seasonality.
F4: Seasonal period.
F5: First-order autocorrelation.
F6: Optimal box-cox transformation parameter.

Visualization

Screen Shot 2020-05-04 at 13 02 11

Discussion

Corporacion, Recruit, Walmart, Rossmann data can be said as 'similar'.
Wikipedia data contains high trend.

kiccho1101 commented 4 years ago

Benchmarks

Screen Shot 2020-05-04 at 13 13 20

kiccho1101 commented 4 years ago

Top-scoring solutions

Screen Shot 2020-05-04 at 13 18 13

As the table shows, time series models were mainly used in first two competitions, but later four competitions mostly feature GBDT and NN.