Testing Order Flow Imbalance in Bitcoin Markets

Overview

This project stress-tests a high-frequency trading (HFT) strategy based on Order Flow Imbalance (OFI) across six major cryptocurrency exchanges.

Data Preprocessing

Raw trade data is processed through a three-step pipeline to handle fragmented liquidity:

1. Unit Normalization: Converts Deribit's inverse futures into standard BTC quantities.
2. Winsorization: Clips extreme outliers (1st/99th percentiles) to prevent block trades from distorting the signal.
3. Z-Score Standardization: Normalizes the signal via a rolling window to measure relative deviation from the mean.

Strategy Logic

The strategy translates order flow into trade decisions systematically:

1. Target Construction: Calculates the "Forward Return" over a set horizon (e.g., 1 second).
2. Model Training: Splits data chronologically (40% Train / 60% Test) and fits a Linear Regression to derive a predictive Beta coefficient.
3. Dynamic Thresholding: Applies the Beta to the out-of-sample set. Trades trigger only if the predicted return exceeds a dynamic "J-Threshold" (targeting the top 5% of opportunities).

Conclusion: Performance & Risk

The backtest reveals a high-alpha signal severely undermined by directional risk and scaling limits.

• Drawdown Risk: On OKX, optimized settings generated a Net P&L of $575,712. However, the Maximum Drawdown was -$1.2M.
• Missing Exit Logic: Without a formal stop-loss, the model averaged down into losing trends, ultimately suffering a catastrophic -$62.3M drawdown.
• Scalability Frictions: Real-world network latency inevitably causes slippage, destroying HFT margins. Furthermore, scaling position limits (e.g., from 10 BTC to 100 BTC) hits a strict liquidity ceiling, triggering immediate adverse selection.

Data Notice

The raw tick and order book data used for this analysis are stored in massive .parquet files that exceed GitHub's repository limits. To comply with these size constraints, the underlying data files have been excluded from this repository.