# Learning-loop review — 2026-07-31

57 closed trades reviewed, 56 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?

Overall expectancy is **+0.68% per trade** across **57 closed trades**, with profit factor 1.38 and total P&L +$268. That's positive, but 57 trades is only just past the "suggestive" range (20–50) the rules ask us to be cautious about, and the win-rate Wilson lower bound (39.9% vs raw 52.6%) shows we can't yet rule out a coin flip on win rate alone. Read as: **cautiously holding, not yet a confirmed edge** — worth continuing to track, not worth celebrating.

## 2. Biggest drag (≥20 trades)

Two buckets clear the 20-trade bar besides "all": **orb-v1** (n=33) and the **stop_loss exit-reason bucket** (n=22). (RSI/MA50/relvol/regime "unknown" buckets also show n=36, but "unknown" isn't an actionable condition — it just means the snapshot field was missing — so it can't inform a parameter change.)

**orb-v1 is the biggest real drag**: 33 trades, win rate 36.4% (Wilson low 22.2%), expectancy **-0.22%**, profit factor **0.70**, total P&L **-$176**. It's the only strategy-level bucket both losing money and large enough to act on (still, at n=33 it sits in the "suggestive" 20–50 zone, so treat conclusions as directional, not final).

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

This is an **entry/stop problem, not an exit problem**, based on the evidence available (note: MAE/MFE aren't broken out per-strategy in this report, so this is inferred from the aggregate stats plus the exit-reason cut, not an orb-specific MAE/MFE table):

- **avgMfeLosersPct is only 0.53%** overall — per the doc's own rule, this is *not* "materially positive." Losers are not running up meaningfully in our favor before reversing, so this isn't a case of "the trade worked and we gave it back." That rules out a take-profit/exit fix.
- **MAE winners (-0.95%) vs MAE losers (-1.55%)** — losers dig only moderately deeper than winners before failing, not the "winner was down big and recovered" pattern that would flag over-tight stops on noisy names.
- The **stop_loss exit bucket** (n=22) is 0% win rate, expectancy -1.82%, average hold only 0.7 days — consistent with orb-v1's own 0.2-day average hold. This looks like trades entering on marginal/false breakout signals that reverse and stop out almost immediately, rather than good signals that later got given back.
- Regime can't be isolated for orb-v1 specifically — the regime cut only shows SPY>MA200 (21 trades, positive expectancy) vs "unknown" (36 trades), with no strategy tag, so this data doesn't support a regime-based read for orb-v1.

Net: this reads as an **entry-confirmation problem** — orb-v1 is taking breakouts that don't have enough follow-through.

## 4. Proposed parameter change

**orb-v1 · `volumeConfirmMult`: 2.5 → 3.0**

Reasoning: the exit and MFE data point away from an exit fix (losers aren't giving back meaningful open profit) and toward weak entry confirmation on fast-failing breakouts (0.2d avg hold, 36.4% win rate, PF 0.70). Raising the volume-confirmation threshold demands a stronger breakout before entry, which should filter out the marginal/false breakouts driving the stop-loss cluster, without touching any other parameter. This is a single, small, reversible change — measure it against orb-v1 trades made *after* the change, not against this historical batch.

*(If no bucket had cleared 20 trades, the correct move would have been NO CHANGE — but orb-v1 at n=33 does clear the bar, so a proposal is warranted here, held to "suggestive" confidence.)*

## 5. Notable open positions (situational color only)

- **qsr has ~50 open tranches**, several visibly underwater: **NXPI** shows 4 open positions all at **-7.31%** — this echoes NXPI's already-poor closed-trade history (qsr·NXPI: 2 trades, -7.93% expectancy, -$109). Same story appears to be repeating live. **TXN** is at -4.23% (2 tranches), **VST** -1.87%, **APP** -1.24%, **ETR** -0.75%, all still open with no calendar time-stop (`maxHoldDays: null` for qsr) — several of these have now been open ~9+ days.
- **MARA** is open in orb-v1 at -2.58% unrealized — orb-v1's only closed MARA trade was also a loss (-3.49% expectancy, -$77), so this may be another repeat pattern to watch once it closes.
- **crypto-trend-v1** (BTC/USD, ETH/USD) has two positions open 10.2 days with **no calendar time-stop** and no closed-trade history in this journal at all — there's no evidence base yet for this strategy; worth watching since it's currently unbounded in hold time.
- None of this should feed into the point-4 parameter proposal — these are open, unresolved trades with no confirmed outcome.