bybit-order-book

SKILL.md

---
name: bybit-orderbook-backtester
description: >
  Download, process, and backtest ByBit derivatives historical order book data. Use this skill when
  the user wants to: (1) download historical order book snapshots from ByBit's derivatives history-data
  page using Selenium automation, (2) process/unzip ob500 JSONL files and filter to depth 50, (3) run
  any of 10 order-book-based trading strategies (Order Book Imbalance, Breakout, False Breakout,
  Scalping, Momentum, Reversal, Spoofing Detection, Optimal Execution, Market Making, Latency
  Arbitrage) against the data, or (4) generate full backtest performance reports with PnL, Sharpe
  ratio, win rate, max drawdown, and strategy comparison. Triggers on: "bybit order book", "order book
  backtest", "download bybit data", "ob500", "order book imbalance", "spoofing detection strategy",
  "market making backtest", "crypto order book", "depth of book backtest", "bybit historical data".
---

# ByBit Order Book Backtester

End-to-end pipeline: download → process → backtest → report.

## Dependencies

```bash
pip install undetected-chromedriver selenium pandas numpy pyarrow --break-system-packages
```

Chrome/Chromium must be installed for Selenium.

## Workflow

The pipeline has 3 stages. Run them sequentially, or skip to later stages if data is already available.

### Stage 1: Download Order Book Data

Prompt the user for:
- **Symbol** (default: BTCUSDT)
- **Date range** (default: last 30 days)

Run `scripts/download_orderbook.py`:

```bash
python scripts/download_orderbook.py \
  --symbol BTCUSDT \
  --start 2024-06-01 --end 2024-06-30 \
  --output ./data/raw
```

Key details:
- Downloads from `https://www.bybit.com/derivatives/en/history-data`
- Automatically chunks into 7-day windows (ByBit's limit)
- Uses `undetected-chromedriver` for Cloudflare bypass
- Outputs: ZIP files in `./data/raw/` named `{date}_{symbol}_ob500.data.zip`
- For data format details: see `references/bybit_data_format.md`

**If Selenium fails** (Cloudflare blocks, UI changes): Instruct the user to manually download from the ByBit page and place ZIPs in `./data/raw/`.

### Stage 2: Process & Filter to Depth 50

Run `scripts/process_orderbook.py`:

```bash
python scripts/process_orderbook.py \
  --input ./data/raw \
  --output ./data/processed \
  --depth 50 \
  --sample-interval 1s
```

What it does:
- Reads JSONL from ZIPs (each line = full 500-level L2 snapshot)
- Filters to top 50 bid/ask levels
- Computes derived features: mid_price, spread, volume_imbalance, microprice
- Optionally downsamples (e.g., `1s`, `5s`, `1min`) — recommended for faster backtests
- Outputs: Parquet files in `./data/processed/`

**Without downsampling**: ~860K snapshots/day, ~300 MB Parquet per day per symbol.
**With 1s downsampling**: ~86K snapshots/day, ~5 MB per day — much more practical.

### Stage 3: Backtest Strategies

Run `scripts/backtest.py`:

```bash
# Run all 10 strategies
python scripts/backtest.py \
  --input ./data/processed/BTCUSDT_ob50.parquet \
  --output ./reports

# Run specific strategies
python scripts/backtest.py \
  --input ./data/processed/BTCUSDT_ob50.parquet \
  --strategies imbalance,breakout,market_making \
  --output ./reports

# Quick test with limited rows
python scripts/backtest.py \
  --input ./data/processed/BTCUSDT_ob50.parquet \
  --max-rows 100000 \
  --output ./reports
```

Strategy keys: `imbalance`, `breakout`, `false_breakout`, `scalping`, `momentum`, `reversal`, `spoofing`, `optimal_execution`, `market_making`, `latency_arb`

Outputs in `./reports/`:
- `{SYMBOL}_backtest_report.json` — Full results with equity curves
- `{SYMBOL}_backtest_report.md` — Comparison table and detailed metrics

Report metrics per strategy: total trades, winners/losers, win rate, cumulative PnL, Sharpe ratio, max drawdown (absolute and %), avg PnL per trade, avg hold time, profit factor, best/worst trade, equity curve.

For strategy logic and tunable parameters: see `references/strategies.md`

## Customization

To modify strategy parameters, edit the `__init__` method of any strategy class in `scripts/backtest.py`. Each strategy's `self.params` dict contains all tunables.

To add a new strategy:
1. Subclass `Strategy` in `scripts/backtest.py`
2. Implement `on_snapshot(self, row, idx, df)` with entry/exit logic
3. Register in `STRATEGY_MAP`

## Troubleshooting

**Selenium can't load ByBit page**: ByBit uses Cloudflare. Ensure `undetected-chromedriver` is up to date. Try `--no-headless` to debug visually. Fall back to manual download.

**Out of memory on processing**: Use `--sample-interval 1s` or larger. Process one day at a time.

**No trades generated**: Strategy thresholds may be too tight for the data period. Relax parameters (lower thresholds, shorter lookbacks) in `references/strategies.md`.