Project · Machine Learning

Crypto Volume Forecasting

A supervised ML pipeline that forecasts 7-day-ahead trading volume for Bitcoin, Ethereum, and Litecoin. Models are trained offline on 730 days of daily OHLCV data from Yahoo Finance using a chronological 80/20 split with a 7-day horizon gap to prevent target leakage. Feature engineering produces ~45 candidate predictors: price and volume lags, rolling statistics, realised volatility, momentum, and calendar effects. A two-pass importance-based pruning step selects the top 20 per model.

Select an asset and model below. Inference loads the pre-trained artifact and runs a live prediction against the most recent market data in under a second.


Run a Forecast


Autoregressive Rolling Forecast

A separate T+1 model predicts next-day volume for the selected asset and model class. Each prediction is fed back as a lag feature to generate T+2, T+3, and so on. Price is held constant across the rollout; only volume updates autoregressively. Forecast error compounds with each step.


Pipeline Details

Data

730 days · BTC, ETH, LTC

Daily OHLCV via Yahoo Finance

Models

Ridge · XGBoost · LightGBM

Trained offline · inference on demand

Target

log-volume, 7 days ahead

Inverted to real scale at inference

Split

80% train · 7-day gap · 20% test

Gap prevents horizon leakage

Features

Top 20 of 45 engineered

Two-pass importance pruning per model

Evaluation

RMSE · MAE · MAPE · R² · Dir. Acc.

Held-out test set · volatility regime