Add docs/methodology.md

docs/methodology.md

@@ -0,0 +1,74 @@
+# Methodology
+
+How the dataset was generated and what biases you should know about.
+
+## The signal
+
+The crash-recovery bot enters when, **on a Polymarket binary or multi-outcome market**:
+
+1. The price has dropped > **N%** from a recent local-window high (`pre_crash_high`).
+2. The current `entry_price` is in a sweet-spot range (default: $0.04 – $0.30 in earlier versions; raised to $0.04 – $0.60 in the current version).
+3. The market is not in a per-token loss cooldown (post-TIMEOUT 7d block).
+4. Market category is not in the persistent blacklist (sports markets where the team has lost the underlying game, etc.).
+5. The orderbook has enough depth at the bid stack to absorb the position size within the slippage budget.
+
+When all conditions hit, the bot opens a position with `size_usd = 5` (standard size in this dataset).
+
+## The exits
+
+The bot closes a position when one of:
+
+- **`RECOVERY`** — `exit_price` reaches a target % of `pre_crash_high` (default: 90% of pre-crash). Most common path. Profitable.
+- **`TIMEOUT_48H`** — held for 48 hours without recovering. Bot exits at whatever the bid stack offers. Sometimes profitable (drift), often a small loss.
+- **`TIMEOUT`** — older shorter-window timeout variant from earlier in the dataset. Same logic, shorter window.
+- **`STOP`** — price keeps dropping below a stop level. Rare in this dataset because the bot's stop is loose (the thesis is "crashes mean-revert," so giving the position room is intentional).
+
+## Known biases
+
+### 1. Survivorship in the trigger
+
+This dataset only contains markets where the trigger fired. If you'd used a different threshold (say, 25% drop instead of the bot's default 20%), you'd see different markets. The data does NOT generalize to "all Polymarket crashes" — it generalizes to "Polymarket crashes that fit this specific signal profile."
+
+### 2. Selection in the entry-price band
+
+The bot only enters when `entry_price` is in the configured range. If a market crashes from $0.80 → $0.50, the bot ignores it (above the range). If a market is at $0.02, the bot ignores it (below the floor). The dataset is therefore **heavy in the $0.04–$0.30 band**.
+
+### 3. Theoretical PnL ≠ realized PnL
+
+`pnl_usd` and `is_profitable` are computed from `entry_price` and `exit_price` — what the bot's order tickets said. Actual on-chain fills typically come in slightly worse, especially for `TIMEOUT_48H` exits where the book is thin. See [pnl-truthteller](https://github.com/LuciferForge/pnl-truthteller) for slippage-adjusted analysis.
+
+### 4. Time period
+
+Data covers **March–April 2026**. This includes:
+- A Polymarket V2 migration period (April 28 cutover) where bot was paused for ~6 hours
+- Various political and sports events specific to that window
+- Polygon network conditions specific to that period (gas costs, liquidity)
+
+Don't assume the patterns extrapolate forward indefinitely. Re-run extraction quarterly as the dataset grows.
+
+### 5. One bot, one strategy, one operator
+
+This is data from a single bot run by a single operator. It is **not** a representative sample of all Polymarket activity, all mean-reversion strategies, or all market participants. Treat it as a case study of one specific strategy executed live.
+
+## What's NOT in the data
+
+- **Order-book depth at entry** — would need historical orderbook snapshots, not currently logged.
+- **Market category** — currently must be looked up via the Polymarket gamma-api using `market_id`.
+- **Time-to-resolution at entry** — same; available via gamma-api.
+- **Other concurrent positions** — capital allocation may have constrained which trades fired.
+- **Slippage** — separate tool: [pnl-truthteller](https://github.com/LuciferForge/pnl-truthteller).
+
+## What kind of analysis this dataset is good for
+
+- **Mean-reversion alpha studies** — does crash-recovery actually work? At what drop_pct does it start working? The data has all the inputs.
+- **Time-of-day effects** — `entry_hour_utc` × `is_profitable` reveals the diurnal pattern.
+- **Hold-time tradeoffs** — the win-rate vs hold-hours curve is in here.
+- **Feature-engineering exercises** — if you can predict `is_profitable` better than 80% accuracy from these features, you've found something.
+- **Backtesting frameworks** — this is real labeled data with real prices, suitable for cross-validation.
+
+## What kind of analysis this dataset is NOT good for
+
+- **General Polymarket research** — too narrow a slice (one bot, one signal).
+- **High-frequency studies** — only entry/exit timestamps, not tick-level data.
+- **Slippage modeling** — see [pnl-truthteller](https://github.com/LuciferForge/pnl-truthteller).
+- **Counterfactuals** ("what would a different bot have done?") — only triggered trades are recorded.