Project case study
ML - Finance · 2025
Crypto Options Price Prediction with Gradient Boosting Trees
Time-series-aware regression model that predicts BTC/ETH option trade prices from Binance data using engineered moneyness features and tuned GradientBoostingRegressor.
Highlights
- Engineered log moneyness, signed moneyness, and ITM flags to capture option structure and nonlinearity.
- Used TimeSeriesSplit cross-validation and a custom composite objective (0.7·R² + 0.3·Normalized RMSE) for realistic out-of-sample tuning.
- Achieved test R² = 0.815 and normalized RMSE = 0.570, with permutation importance confirming core pricing drivers.
Gallery
What
Built a regression model to predict cryptocurrency option trade prices using Binance BTC/ETH options transaction data.
Focused on out-of-sample performance and interpretability by modeling how moneyness, time to maturity, and realized volatility relate to observed trade prices.
How
Preprocessed transaction data and encoded option type (call/put) and underlying asset (BTC/ETH), then engineered log moneyness, signed moneyness, and an in-the-money (ITM) flag.
Benchmarked linear regression against tree ensembles, then selected GradientBoostingRegressor for nonlinear structure and stable generalization.
Switched from shuffled KFold to TimeSeriesSplit to avoid overly optimistic validation and tuned hyperparameters with RandomizedSearchCV using a custom composite score (0.7·R² + 0.3·Normalized RMSE).
Validated drivers with permutation importance and generated diagnostic plots including actual vs predicted pricing.
Results
Achieved strong out-of-sample accuracy on the test set with R² = 0.815 and normalized RMSE = 0.570.
Identified moneyness, underlying price, strike price, and realized volatility lags as the dominant drivers of predicted option prices.
Produced a time-series-aware, reproducible modeling pipeline suitable for practical financial forecasting and evaluation.