Reading Market Turbulence

Forecasts short-horizon realized volatility from real crypto order-flow — 24.8 million real trades, BTCUSDT + ETHUSDT, 10 full days each, downloaded directly from Binance's public historical data archive (no API key, no synthetic data anywhere).

Code and method Full README

This page shows the project's results. The methodology, the design decisions and the limitations are documented in the repository's README.

Task

Price and realized volatility over the sample window

Order-flow imbalance vs. next-bucket realized volatility

Results (real run, GroupKFold across 5 real trading days)

Activation comparison with a custom loss (PyTorch)

Activation comparison (ReLU vs. GELU vs. Swish)

Fresh confirmation run (real data, re-executed today)

Predicted vs. actual realized volatility, held-out day, fresh run — **Interactive**: actual vs. predicted next-bucket realized volatility over the held-out day, with VWAP overlaid — open the interactive chart (self-contained HTML, Plotly, pan/zoom/hover).

LightGBM feature importance, fresh run — **Interactive**: actual vs. predicted next-bucket realized volatility over the held-out day, with VWAP overlaid — open the interactive chart (self-contained HTML, Plotly, pan/zoom/hover).