# Learning-loop review — 2026-07-26

23 closed trades reviewed, 14 currently open. Proposal only — nothing here is applied automatically (CLAUDE.md §4: the AI review proposes, dan disposes).

# AI Review — Paper Trading Journal

## 1. Is overall expectancy holding?

No — overall expectancy is **-0.27%** per trade across 23 closed trades, with profit factor 0.56 (well under 1.0, i.e. losing money on a gross basis). The Wilson lower bound on win rate is only 25.6% against a raw 43.5%, so we can't even confidently claim the win rate is above "mediocre coin flip" territory. At n=23 this sits just barely past the ~20-trade floor — per the rules this is "suggestive only," not a confirmed result. Directionally the number is bad, but treat it as a hypothesis, not a verdict.

## 2. Which bucket is the biggest drag?

Two buckets clear the ~20-trade minimum: **all trades (n=23)** and **orb-v1 (n=20)**. No other cut (RSI, MA50 distance, rel-volume, regime, exit reason, symbol, strategy+symbol) reaches 20, so none of those may influence a decision yet.

Of the two qualifying buckets, **orb-v1 is the drag**: 20 trades, win rate 35.0% (Wilson low 18.1%), expectancy **-0.52%**, profit factor 0.43, total P&L **-$248**. It accounts for essentially all of the portfolio's loss (vwap-mr-v1, at only 3 trades, is not a decision-grade sample and is excluded here per rule 5).

## 3. Entry problem, exit problem, or regime problem?

Regime can't be diagnosed at all — the "by market regime" bucket is entirely `unknown` (snapshot data isn't populating that field), so that lens is unavailable this cycle.

Looking at MAE/MFE (only available at the overall level, but orb-v1 is 20 of the 23 trades so it's a reasonable proxy):

- **MAE losers avg -1.30%** — losing trades aren't cratering deeply before stopping out; the adverse move is modest.
- **MFE losers avg 0.46%** — losers weren't running up meaningfully in profit before reversing. This is *not* "materially positive," so the METRICS.md exit-problem signature (winning trade given back) isn't present.
- The exit-reason table reinforces this: **stop_loss (n=12) has a 0% win rate and -1.48% expectancy** — a large share of orb-v1's trades are simply getting stopped, while eod_flatten (n=8, 87.