Review of kaggle time series competition

Inspired by Learnings from Kaggle’s Forecasting Competitions by Casper Solheim Bojer & Jens Peder Meldgaard in 2020, I surveyed the top 3 solutions in the past kaggle time series competitions since 2014 to 2024.

If you find new time series competitions, please tell me by issues.

❤️ Support This Project

If you find this project helpful, please consider supporting it with a small donation!

Table of Contents

• List of competitions
• Top 3 most voted EDAs
  • 1. Walmart Recruiting - Store Sales Forecasting
  • 2. Walmart Recruiting II: Sales in Stormy Weather
  • 3. Rossmann Store Sales
  • 4. Predicting Red Hat Business Value
  • 5. Web Traffic Time Series Forecasting
  • 6. TalkingData AdTracking Fraud Detection Challenge
  • 7. Corporación Favorita Grocery Sales Forecasting
  • 8. Recruit Restaurant Visitor Forecasting
  • 9. Google Analytics Customer Revenue Prediction
  • 10. LANL Earthquake Prediction
  • 11. Two Sigma: Using News to Predict Stock Movements
  • 12. ASHRAE - Great Energy Predictor III
  • 13. University of Liverpool - Ion Switching
  • 14. M5 Forecasting - Accuracy
  • 15. Jane Street Market Prediction
  • 16. Acea Smart Water Analytics
  • 17. Google Brain - Ventilator Pressure Prediction
  • 18. Optiver Realized Volatility Prediction
  • 19. G-Research Crypto Forecasting
  • 20. Ubiquant Market Prediction
  • 21. American Express - Default Prediction
  • 22. GoDaddy - Microbusiness Density Forecasting
• Top 3 solutions
  • 1. Walmart Recruiting - Store Sales Forecasting
  • 2. Walmart Recruiting II: Sales in Stormy Weather
  • 3. Rossmann Store Sales
  • 4. Predicting Red Hat Business Value
  • 5. Web Traffic Time Series Forecasting
  • 6. TalkingData AdTracking Fraud Detection Challenge
  • 7. Corporación Favorita Grocery Sales Forecasting
  • 8. Recruit Restaurant Visitor Forecasting
  • 9. Google Analytics Customer Revenue Prediction
  • 10. LANL Earthquake Prediction
  • 11. Two Sigma: Using News to Predict Stock Movements
  • 12. ASHRAE - Great Energy Predictor III
  • 13. University of Liverpool - Ion Switching
  • 14. M5 Forecasting - Accuracy
  • 15. Jane Street Market Prediction
  • 16. Acea Smart Water Analytics
  • 17. Google Brain - Ventilator Pressure Prediction
  • 18. Optiver Realized Volatility Prediction
  • 19. G-Research Crypto Forecasting
  • 20. Ubiquant Market Prediction
  • 21. American Express - Default Prediction
  • 22. GoDaddy - Microbusiness Density Forecasting

List of competitions

#	Year	Title	Data size
1	2014	Walmart Recruiting - Store Sales Forecasting	3.22MB
2	2015	Walmart Recruiting II: Sales in Stormy Weather	9MB
3	2015	Rossmann Store Sales	39.85MB
4	2016	Predicting Red Hat Business Value	26.74MB
5	2017	Web Traffic Time Series Forecasting	611.85MB
6	2018	TalkingData AdTracking Fraud Detection Challenge	11.27GB
7	2018	Corporación Favorita Grocery Sales Forecasting	479.88MB
8	2018	Recruit Restaurant Visitor Forecasting	27.3MB
9	2018	Google Analytics Customer Revenue Prediction	35.9GB
10	2019	LANL Earthquake Prediction	10.42GB
11	2019	Two Sigma: Using News to Predict Stock Movements	Not available
12	2019	ASHRAE - Great Energy Predictor III	2.61GB
13	2020	University of Liverpool - Ion Switching	146.08MB
14	2020	M5 Forecasting - Accuracy	450.47MB
15	2020-2021	Jane Street Market Prediction	Not available
16	2020-2021	Acea Smart Water Analytics	3.45MB
17	2021	Google Brain - Ventilator Pressure Prediction	698.79MB
18	2022	Optiver Realized Volatility Prediction	2.73GB
19	2022	G-Research Crypto Forecasting	3.12GB
20	2022	Ubiquant Market Prediction	18.55GB
21	2022	American Express - Default Prediction	50.31 GB
22	2022-2023	GoDaddy - Microbusiness Density Forecasting	10.93 MB

Top 3 most voted EDAs

To learn the characteristic of data given in each competition, EDA is one of the best way.
So top 3 most voted EDAs are listed.

1. Walmart Recruiting - Store Sales Forecasting

> Go to the top

1. EDA and Store Sales Predictions using XGB
2. Walmart prediction - (1) EDA with time and space
3. Wallmart Sales - EDA - feat eng [Future Update]

2. Walmart Recruiting II: Sales in Stormy Weather

> Go to the top

NA

3. Rossmann Store Sales

> Go to the top

1. Time Series Analysis and Forecasts with Prophet
2. EDA and forecasting with RFRegressor_FINAL_UPDATED
3. How Does New Competition Affect Sales?

4. Predicting Red Hat Business Value

> Go to the top

1. Time Travel (EDA)
2. Redhat EDA
3. RedHat Hack in plain English (EDA)

5. Web Traffic Time Series Forecasting

> Go to the top

1. Wiki Traffic Forecast Exploration - WTF EDA
2. Web Traffic Time Series Forecasting (EDA)
3. Wikipedia Web traffic EDA

6. TalkingData AdTracking Fraud Detection Challenge

> Go to the top

1. TalkingData EDA plus time patterns
2. TalkingData EDA and Class Imbalance
3. TalkingData: EDA to Model Evaluation | LB: 0.9683

7. Corporación Favorita Grocery Sales Forecasting

> Go to the top

1. Shopping for Insights - Favorita EDA
2. Memory optimization and EDA on entire dataset
3. Grocery EDA Dirty XGBoost, Arima,ETS,Prophet

8. Recruit Restaurant Visitor Forecasting

> Go to the top

1. Be my guest - Recruit Restaurant EDA
2. Exhaustive Weather EDA/File Overview
3. Recruit Restaurant EDA

9. Google Analytics Customer Revenue Prediction

> Go to the top

1. R EDA for GStore + GLM + KERAS + XGB
2. Google Analytics EDA + LightGBM + Screenshots
3. A Very Extensive GStore Exploratory Analysis

10. LANL Earthquake Prediction

> Go to the top

1. Earthquakes FE. More features and samples
2. LANL Earthquake EDA and Prediction
3. Masters Final Project: EDA

11. Two Sigma: Using News to Predict Stock Movements

> Go to the top

1. EDA, feature engineering and everything
2. 👨‍🔬 Bird Eye 👀 view of Two Sigma + NN Approach
3. Simple EDA - Two Sigma

12. ASHRAE - Great Energy Predictor III

> Go to the top

1. 🔌⚡ASHRAE -Start Here: A GENTLE Introduction
2. EDA for ASHRAE
3. A deep dive EDA into ALL variables

13. University of Liverpool - Ion Switching

> Go to the top

1. Ion Switching Competition : Signal EDA 🧪
2. EDA - Ion Switching
3. Simple EDA-Model

14. M5 Forecasting - Accuracy

> Go to the top

1. Back to (predict) the future - Interactive M5 EDA
2. M5 Competition : EDA + Models 📈
3. Time Series Forecasting-EDA, FE & Modelling📈

15. Jane Street Market Prediction

> Go to the top

1. Jane Street: EDA of day 0 and feature importance
2. Jane_street_Extensive_EDA & PCA starter 📊⚡
3. EDA / A Quant's Prespective

16. Acea Smart Water Analytics

> Go to the top

1. Acea Smart Water: Full EDA & Prediction
2. EDA: Quenching the Thirst for Insights
3. Quick EDA | Reporting & Data Understanding

17. Google Brain - Ventilator Pressure Prediction

> Go to the top

1. Ventilator Pressure Prediction: EDA, FE and models
2. 🔥EDA +FE+TabNet 🧠🧠[Weights and Biases]
3. Ventilator Pressure: EDA and simple submission

18. Optiver Realized Volatility Prediction

> Go to the top

1. Optiver Realized: EDA for starter(English version)
2. Optiver Realized Volatility Prediction - EDA
3. Optiver; EDA XGBoost starter(日本語,Japanese)

19. G-Research Crypto Forecasting

> Go to the top

1. 📊 G-Research Plots + EDA 📊
2. To The Moon 🚀 [G-Research Crypto Forecasting EDA]
3. 📈📊[G-crypto] Interactive Dashboard + Indicators

20. Ubiquant Market Prediction

> Go to the top

1. EDA- target analysis
2. Ubiquant EDA and Baseline
3. 🔥The most advanced analytics🔥

21. American Express - Default Prediction

> Go to the top

1. AMEX EDA which makes sense ⭐️⭐️⭐️⭐️⭐️
2. AMEX Default Prediction EDA & LGBM Baseline
3. American Express EDA

22. GoDaddy - Microbusiness Density Forecasting

> Go to the top

1. TBD
2. TBD
3. TBD

Top 3 solutions

1. Walmart Recruiting - Store Sales Forecasting

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1	Time Series Models: • STLF/ETS • STLF/ARIMA • Seasonal ARIMA • Fourier ARIMA	• SVD preprocessing • Time series decomposition • Correlation-based NN pooling • Holiday period adjustment	Average of 6 time series models: • 3 simple models • 3 advanced models	Department-wise pooling across stores	💻	🔊
2	Statistical Methods: • ARIMA with STL decomposition • ETS with STL decomposition • Naive method with STL decomposition • UCM Machine Learning Methods: • Random Forest • Linear Regression • K Nearest Regression • Principal Component Regression	• Used week of the year (1-52) • Filled missing values with 0 • Different holiday weighting for stores with high growth rate vs. stores without high growth	Weighted average and trimmed mean of 6 methods*	Individual models for each store- department combination (~3600 total)	💻	🔊
3	Simple year-over-year approach (no statistical or ML models): • Prior year sales as base predictor • Holiday week alignment • Weighted average of prior weeks • Store & department trend factors • Manual trend coefficients	• Calendar date matching between years • Special handling for moving holidays • Store-specific trend factors • Department-specific trend factors • Minimal "warm day" adjustment	None - single model with minor adjustments	Individual predictions for each store- department combination	💻	🔊

2. Walmart Recruiting II: Sales in Stormy Weather

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1	PPR (Projection Pursuit Regression) for curve fitting + Lasso regression with vowpal wabbit	• Weekday, weekend flags, holiday flags • Item/store numbers • Date features (year, month, day) • Black Friday period flags • Weather features (precipitation>0.2, temperature departure>8 or <-8) • Moving average • Handling of consecutive zeros • Feature interactions (AB, AC, BM, CM, BK, CK)	Single model	• Excluded item/stores with all zeros • Excluded data from 2013-12-25 • Excluded data where moving average was zero • Excluded dates with too many consecutive zeros on both sides	💻	🔊
2	-	-	-	-	NA	NA
3	-	-	-	-	NA	NA

3. Rossmann Store Sales

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1					NA	🔊
2					NA	NA
3					💻	🔊

4. Predicting Red Hat Business Value

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1	Two-level approach: • 1st level: XGBoost classifier on aggregated data • 2nd level: XGBoost with leakage exploitation	• TF-IDF features for categorical variables • Group level aggregations • Group_1 ID value • Number of activities/people in group • Min/max dates • First/last activity in timeline	Average of: • Best LB model • Best CV model	• Removed group_1=17304 (30% of train data) • Used Distinct operator for group_1's with 3000+ rows • 5-fold unstratified CV based on people file	NA	🔊
2	Mixture of three models: • Logistic regression • kNN • XGBoost (based on public scripts) + Probabilistic interpolation model for groups	• Histogram-based features (fuzzy binary encoding) • Date-based probabilistic interpolation • Leaderboard feedback	Weighted average: • 0.4 (linear model) • 0.25 (kNN) • 0.25 (public script) • 0.1 (0.5 constant)	NA	NA	🔊
3	XGBoost + extensive group_1 and date trick exploitation	• Modified leave-one-out encoding for char_10 • Group_1 as continuous feature • Time-based confidence levels for leak exploitation • Group-person level leak simulation for training data	Post-processing: • 90% weight to most extreme prediction within a group • 10% weight to raw prediction • Rule-based overrides for ML shortcomings	NA	NA	🔊

5. Web Traffic Time Series Forecasting

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1					💻	🔊
2					💻	🔊
3					💻	🔊

6. TalkingData AdTracking Fraud Detection Challenge

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1					NA	🔊
2					NA	🔊
3					NA	🔊

7. Corporación Favorita Grocery Sales Forecasting

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1					💻 💻	🔊
2					NA	🔊
3					NA	🔊

8. Recruit Restaurant Visitor Forecasting

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1	LightGBM		-	-	💻 💻	NA
2	-		-	-	NA	NA
3	-		-	-	NA	NA

9. Google Analytics Customer Revenue Prediction

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1					💻	🔊
2					NA	🔊
3	-		-	-	NA	NA

10. LANL Earthquake Prediction

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1					💻 💻	🔊
2					NA	🔊
3					NA	🔊

11. Two Sigma: Using News to Predict Stock Movements

> Go to the top

NA

12. ASHRAE - Great Energy Predictor III

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1	CatBoost LightGBM MLP				NA	🔊
2	XGBoost LightGBM Catboost Feed-forward Neural Network				NA	🔊
3	CNN LightGBM Catboost				NA	🔊

13. University of Liverpool - Ion Switching

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1					NA	🔊
2					💻	🔊 🔊
3					NA	🔊 🔊

14. M5 Forecasting - Accuracy

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1	LightGBM				NA	💻
2	LightGBM				NA	💻
3	DeepAR				NA	💻

15. Jane Street Market Prediction

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1	XGBoost NN				💻	NA
3	49 layers MLPs	No	15 ensembles of NN		NA	🔊

NA for Pos #2

16. Acea Smart Water Analytics

> Go to the top

NA

17. Google Brain - Ventilator Pressure Prediction

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1	LSTM Transformer		single architecture	KFold	💻	🔊
2	Stacked LSTM		ensembled by 7 models	KFold	NA	🔊
3	Conv1d Stacked LSTM		random seed average	Stratified K-Folds	NA	🔊

18. Optiver Realized Volatility Prediction

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1	LightGBM MLP CNN		equally weighd average	GroupKFold	💻	🔊
3	LightGBM MLP TabNet		equally weighd average	KFold	💻	🔊

NA for Pos #2

19. G-Research Crypto Forecasting

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1	-	-	-	-	NA	NA
2	LightGBM		Single model		NA	🔊
3	LightGBM		Single model		💻 💻	🔊

20. Ubiquant Market Prediction

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1	LightGBM TABNET		Average of (LGBM x 5 Folds) + (TABNET x 5 Folds)	PurgedGroupTimeSeries TimeSerieseSplit KFold	NA	🔊
2	LightGBM		-	Purged K-FOLD cross validation with embargo	NA	🔊
3	6 layers transformer		5 seeds ensemble	-	NA	🔊

21. American Express - Default Prediction

> Go to the top

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1	LightGBM GRU		Ensembled by 4 models		💻	🔊
2	LGB/XGB/CTB NN				NA	🔊
3	LGB/CTB		Ensembled by 3 models		NA	🔊

22. GoDaddy - Microbusiness Density Forecasting

> Go to the top

Ongoing

Pos	Methods	FE	Ensemble	Split	Code	Discussion
1	Linear regression				💻	🔊
2
3