# Learning-loop review — 2026-08-06

106 closed trades reviewed, 83 currently open. Proposal only — nothing here is applied automatically (CLAUDE.md §4: the AI review proposes, dan disposes).

## 1. Is overall expectancy holding?

Overall expectancy is **+1.41% per trade** across **106 closed trades**, with profit factor 1.93 and a Wilson-lower-bound win rate of 56.6% (raw 66.0%). The overall sample clears the ~20-trade minimum, so this headline number is meaningful — expectancy looks like it's holding, and it's not just a coin-flip win rate propped up by a few outliers.

## 2. Which bucket is the biggest drag?

Restricting to buckets with ≥20 trades: `orb-v1` (39), `qsr` (52), `rsi 50-70` (21), `relvol<0.8` (36), `SPY>MA200` (40), `stop_loss` exit (30), `take_profit` exit (39), and `unknown` (66, in several cuts).

Of these, **`orb-v1` (strategy) is the clear drag**: n=39, win rate 41.0% (Wilson low 27.1%), expectancy **-0.08%**, profit factor **0.85**, total P&L **-$102**. It's the only strategy-level bucket that's actually losing money with a sample large enough to act on.

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

The report doesn't break MAE/MFE out by strategy — only aggregate figures are given (MAE winners -1.79%, MAE losers -3.14%, MFE losers +0.82%). That aggregate MFE-losers number is only mildly positive, well short of the "materially positive" threshold the doc flags as an exit/take-profit problem, and it's pooled across qsr/vwap/orb so it can't be pinned on orb-v1 specifically. I don't have the strategy-specific MAE/MFE split needed to make a clean call here — that's a data gap, not something to guess past.

What the strategy-level numbers *do* support: orb-v1's Wilson lower bound (27.1%) sits well below its raw win rate (41.0%) — we can't rule out this being closer to a coin flip. Combined with a very short avg hold (0.4d) and payoff ratio of only 1.29 (wins aren't big enough to offset the low hit rate), this looks more consistent with an **entry problem** (too many false breakouts getting stopped out quickly) than a "we let a winner round-trip" exit problem — but this is a read from win-rate/payoff shape, not from a confirmed MAE pattern, since that data isn't available at the strategy level. Not a regime call either: `SPY>MA200` bucket (n=40) is actually healthy (expectancy +1.50%, PF 2.78), so there's no evidence the drag is regime-driven.

## 4. Proposed parameter change

**Strategy: `orb-v1` — `volumeConfirmMult`: 2.5 → 3.0**

Reasoning: orb-v1's failure mode (41% win rate, Wilson-low 27%, 0.4-day average hold, PF 0.85) is consistent with the breakout signal firing on volume that isn't actually confirming real follow-through — i.e., false breakouts. Per METRICS.md's own prescription for this pattern ("enter later / demand more confirmation"), tightening the volume-confirmation multiplier is the smallest, most direct lever to filter weaker signals without touching sizing, targets, or the stop logic simultaneously. This is one parameter, on the one bucket that both clears the 20-trade bar and shows negative expectancy.

I'm explicitly *not* touching `targetRMultiple` or `entryCutoffMinute` at the same time — only this one knob, to keep attribution clean per rule #2.

## 5. Notable about open positions (context only, not evidence for #4)

- The open book is almost entirely **qsr** (single strategy carrying ~40+ open legs), and `qsr` has **`maxHoldDays: null`** — several positions (AEM, HDB, TXN) have been open **15.3 days** with no calendar time-stop forcing a decision either way.
- **VST is the most underwater open position at -6.08%** unrealized, sitting well past its stop distance visually but apparently not yet triggered — worth a manual look.
- TXN (-3.23%) and MO (-1.64%) are also lagging without resolution.
- The two crypto-trend-v1 positions (BTC/USD, ETH/USD) have been open **16.2 days**, also with no calendar time-stop — that strategy isn't even in the four under review here, but it's the longest-aged position in the book and worth someone's attention independent of this analysis.
- None of this open-position color should feed into the orb-v1 parameter decision above — those trades have no confirmed outcome yet (rule: judge only on closed, post-change trades).