Optiver Trading Competition

Overview

This project performs feature engineering and modeling on order book time series data to evaluate the effectiveness of Technical Analysis features in light of predicting the target

The main tasks include:

• Exploratory data analysis
• Feature engineering
  • Technical indicators (RSI, MACD, Bollinger Bands)
  • Time-weighted order book features
• Feature selection
  • Correlation analysis
  • Filter-based (ANOVA)
  • Wrapper-based (RFE)
• Modeling
  • XGBoost
  • Hyperparameter tuning
  • Cross-validation

Data

The train.csv dataset contains the following key features:

• stock_id: Stock identifier
• bid/ask_price: Bid and ask prices in the order book
• bid/ask_size: Bid and ask sizes in the order book
• target: The target variable to predict

Contents

Data Exploration and Preprocessing

• Handle missing values
• Feature scaling

Feature Engineering

• Technical indicators
  • RSI
  • MACD
  • Bollinger Bands
• Time-weighted order book features

Feature Selection

• Correlation analysis
• ANOVA feature selection
• Recursive feature elimination

Modeling

• XGBoost regressor
• Hyperparameter tuning with Optuna
• Cross-validation

Installation

This project requires Python 3 and the following libraries:

• Pandas
• NumPy
• Scikit-learn
• XGBoost
• Optuna

These can be installed with pip or conda.

Results

We evaluated the impact of adding technical analysis (TA) features to our orderbook model across different numbers of stocks: 10, 50, and 100.

The key hypotheses tested were:

1. TA features reduce MAE by at least 10% compared to no TA features.
2. TA features reduce MAE compared to no TA features.

Hypothesis 1 Results

H0: TA features reduce MAE by at least 10%
H1: TA features reduce MAE by <10%

# Stocks	Result
10	✘
50	✘
100	✓

• For 10 and 50 stocks, the >10% MAE improvement null hypothesis is rejected
• For 100 stocks, the null >10% MAE improvement hypothesis holds

Hypothesis 2 Results

H0: TA features reduce MAE
H1: TA features do not reduce MAE

# Stocks	Result
10	✓
50	✓
100	✓

• For all stock counts, TA features lead to lower MAE
• The alternative hypotheses are rejected

Conclusion: Adding TA features reduces model MAE compared to no TA features for all stock counts. However, the >10% MAE improvement only holds for 100 stocks based on the Welch test results.