Historian

The historian reads the bet log and summarizes how the bot has performed over time. It runs after analysis each window and powers the weekly review that drives the agentic loop.

Data source

Every pipeline run appends one row per window to the bet log. Rows include:

• Trade decision (UP, DOWN, or NO_TRADE)
• Score, confidence, EV, setup grade
• Full signal snapshot (pattern, trends, indicator biases, score breakdown)
• Gate audit fields when the judge rejected a trade
• Outcome and PnL after market resolution

The historian filters out phantom audit rows and aggregates the rest.

What it computes

Headline stats

• Total bets, resolved vs open, win rate, total PnL
• Daily bet count and daily PnL (UTC)
• Average confidence and average EV across resolved trades
• Profit factor and consecutive loss streak

Rolling windows

Recency-weighted stats over the last 10 and 20 resolved bets: win rate, PnL, profit factor, average EV. Recent bets count more via an exponential half-life (default 35 bets).

Pattern taxonomy

The historian groups bets by labels and tracks win rate per group:

• Core: side, grade, mode
• Setup: pattern name, trend alignment
• Market: regime, liquidity grade, spread band
• RSI: oversold / overbought / neutral zones
• Combo: pattern + trend + regime combinations

A pattern needs at least 8 samples before it appears in stats. Combo labels need enough data to be reliable.

Drawdown and session momentum

• Peak-to-trough drawdown (absolute and percentage)
• Recovery time from the last peak
• Session window stats (default 2 hours): bets, win rate, PnL, hot/cold flags

EV calibration

Compares predicted EV from the judge against realized outcomes. Flags overconfidence when predicted edge consistently overshoots actual results. Calibration data also feeds dirProb blending in the judge once sample size is sufficient.

Runtime vs weekly review

On each pipeline run the historian produces a status snapshot (win rate, streak, pattern highlights). That snapshot is logged for monitoring.

The deeper work happens in the weekly profitability review:

1. Map all judge gates and sizing formula
2. Analyze live trades vs skipped windows
3. Review bet sizing and historian calibration
4. Output ranked recommendations in a final report

Agents read that report and propose changes to weights, gates, sizing, and risk limits. Approved updates redeploy to the live bot through the agentic loop.

Why it matters

Without the historian, gate thresholds would stay static. With it, Catalysm can:

• Block patterns that lose across every regime
• Tighten or loosen dead zones based on skip analysis
• Adjust grade multipliers when A-grade setups underperform
• Track whether predicted EV matches reality

See stats methodology for how public-facing win rates on the Bots page relate to the same bet log.