# Learning-loop review — 2026-08-11

157 closed trades reviewed, 140 currently open. Proposal only — nothing here is applied automatically (CLAUDE.md §4: the AI review proposes, dan disposes).

## 1. Is overall expectancy holding?

Pooled expectancy is **+1.69% per trade** over **157 closed trades**, profit factor **1.98**, max drawdown only **$287**. The pooled sample clears the ~20-trade minimum, so this headline number is meaningful *as a pooled statistic* — but recall it mixes four strategies with very different trade counts and two different account risk configs (paper-main has 142 of the 157 trades), so "overall expectancy" here is really "paper-main + qsr expectancy," not a clean single-strategy read.

## 2. Biggest drag bucket (≥20 trades)

**orb-v1**, by strategy: n=46, expectancy **-0.06%**, profit factor **0.87**, win rate 41.3% (Wilson lower bound 28.3%). It's the only strategy-level bucket that is net negative and clears the 20-trade bar, and it's dragging the overall scoreboard down ($-91 total P&L over 46 trades).

(The `stop_loss` exit-reason bucket, n=41, expectancy -3.77%, is even more negative in isolation, but that's a tautological bucket — stop-loss exits are losses by construction, pooled across all strategies — so it's not informative as "the drag," it's just the shape of the loss side of the distribution.)

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

The report does **not** give MAE/MFE broken out by strategy — only the pooled figures (MAE winners -1.88%, MAE losers -3.53%, MFE losers 0.83%), which are dominated by qsr (95 of 157 trades) and can't be trusted to describe orb-v1's specific failure mode. I can't cleanly cite an orb-v1-specific MAE/MFE pattern, so I won't force one.

What the strategy-level numbers *do* show: orb-v1's payoff ratio is **1.31** (average win bigger than average loss) while win rate is only **41.3%**. Low win rate + reasonable payoff points more toward an **entry-quality problem** (too many breakouts triggering that don't follow through) rather than an exit/target problem (which would show up as payoff <1, i.e. giving back winners). No regime cut exists for orb-v1 specifically either, so regime can't be ruled in or out from this data. Net: best-supported explanation is entry quality, but this is a hypothesis, not a firm MAE-backed conclusion — a strategy-crossed-with-MAE/MFE cut would be needed to confirm.

## 4. Proposed parameter change

**orb-v1.volumeConfirmMult: 2.5 → 3.0**

Reasoning: orb-v1's problem looks like too many marginal breakouts getting triggered (low win rate, decent payoff when they do work). `volumeConfirmMult` is the single lever that raises the bar on breakout conviction without touching target/stop/hold logic, so it isolates the "entry quality" hypothesis from question 3. This is a single, small, reversible change — no other orb-v1 or other-strategy parameters touched. It should be judged only on orb-v1 trades taken *after* the change, against another ~20+ trade sample before acting further (rule 4).

If no bucket had cleared 20 trades I would have proposed no change — but orb-v1 does clear it, so a change is proposable here.

## 5. Notable about open positions

- QSR dominates open exposure (~130+ open QSR legs) and `maxHoldDays` is `null` for qsr — there is no calendar time-stop, and several positions have been open 12–20+ days sitting at small persistent losses (HDB -2.58% at 20.3d, MO -4.58% at 5.2d, multiple ETR/AZN/CI legs around -0.9% to -1.8% for 4–12 days). None are catastrophic, but it's a lot of aging, unresolved risk with no time-based exit mechanism.
- The two crypto-trend-v1 positions (BTC/USD, ETH/USD) have been open 21.2 days with no unrealized P&L reported at all — worth someone checking that P&L is actually being tracked for that pair.
- This is color only — none of it feeds into the point-4 proposal, since none of these trades have a confirmed outcome yet.

## 6. QSR shadow comparison note

Worth flagging qualitatively: the legacy (isTriggered+isBuy/buyZonePct) shadow method and the live newlyA method are picking **almost entirely different tickers** — only AEP and SU overlap out of 8 tickers each side (legacy: AEP, DDOG, IBKR, MO, AZN, SU, COHR, MAR; live: AEP, SU, TRP, BTI, EBAY, HON, ITUB, NKE). There's also a large frequency gap: legacy fired 16 would-buy signals over the window vs. the live method producing 48 open entries (0 closed) on its 8 tickers — the live method appears to be firing considerably more often, though "more open positions" isn't a clean apples-to-apples signal-count comparison since one entry can spawn multiple tranches. The shadow window is still running (ends 2026-08-23), so no closed-window outcome comparison is available yet — once it closes, a fuller backtest comparing the two selection methods' actual (simulated) outcomes could be requested. This is observational only and is explicitly excluded from the point-4 proposal per the source data's own caveat.