# Learning-loop review — 2026-08-13

193 closed trades reviewed, 144 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 **+1.70% per trade** across 193 closed trades — positive, and the pooled n clears the ~20-trade bar. But two caveats matter before trusting that number at face value:

- It's dominated by **qsr** (125 of 193 trades, 65%), so "overall" is really "mostly qsr's performance."
- Per the account breakdown, 172 of the 193 trades are in **paper-main** (simulated, optimistic fills). The two smaller accounts (`Paper02qsr`, n=9; `live-1`, n=12) are both well under the 20-trade minimum, so neither can independently confirm the edge yet, and `live-1`'s trades are price-biased (small notional cap) rather than a clean replicate.

So: yes, expectancy is holding and positive, but the evidence base is really "qsr in paper, at reasonable sample size" — not yet confirmed live or across accounts.

## 2. Which bucket is the biggest drag?

Restricting to buckets with **n ≥ 20**: the clear drag is the **orb-v1 strategy** (n=52, expectancy **-0.11%**, profit factor **0.84**, win rate 42.3%, Wilson lower bound 29.9%). It's the only strategy- or condition-level bucket at ≥20 trades with negative expectancy and PF<1.

(Note: `stop_loss` exit-reason bucket, n=49, shows -3.76% expectancy/0% win rate — but that's tautological, since "exited via stop" trivially means "lost money by definition of the label," not an independent drag signal.)

## 3. Entry (MAE), exit (MFE), or regime problem?

The report only gives MAE/MFE **aggregated across all strategies** (MAE winners -1.73%, MAE losers -3.55%, MFE losers +0.80%), not broken out per strategy — so a clean MAE/MFE diagnosis specific to orb-v1 isn't available from this data, and that limitation should be stated rather than papered over.

That said, the numbers that *are* available point toward an **entry/selectivity problem, not an exit problem**:
- The overall MFE-losers figure (+0.80%) is well below "materially positive" — the doc's own bar for calling it an exit/take-profit issue. There's no aggregate sign of "trade worked, we gave it back."
- orb-v1's own payoff ratio is 1.16, implying a breakeven win rate of ~46%. Its actual win rate is 42.3% (Wilson lower bound 29.9%) — it's losing on selectivity, not on how winners/losers are managed once open. That's consistent with the MAE-side story in the doc ("stop tighter than natural noise" / weak entry confirmation), even though we can't confirm it directly without a strategy-specific MAE split.

## 4. One concrete parameter change

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

**Reasoning:** orb-v1's win rate (42.3%) sits below the breakeven threshold implied by its own payoff ratio (1.16 → need ~46%), and PF is 0.84. That's a signal-quality/entry problem, not an exit problem (MFE-losers isn't elevated at the aggregate level). Raising the volume-confirmation bar is the smallest, most direct lever to filter out marginal breakouts without touching exits, targets, or sizing. This is one parameter, one direction, and should be judged only on orb-v1 trades taken *after* the change (per rule 4), not by rescoring the current 52.

(orb-v1 is backtest-only/not live-wired, so this change carries no live-capital risk to test.)

## 5. Notable in currently open positions

- QSR is running a **large number of open, unresolved positions** (107 in paper-main, 22 in Paper02qsr, 15 in live-1) with `maxHoldDays: null` — no calendar time-stop — and many are aging 7–22+ days across multiple pyramided tranches per ticker (AZN, ETR, ARGX, SNY, CI, BTI, HON, NKE, TRP, AEP, WMB, HDB, TRGP, COIN).
- **HDB** stands out: two positions at **-4.21%** for **22.3 days**, both underwater and aging the longest in the book. This is worth flagging because HDB's *closed* trade history is also poor (qsr·HDB: 2 trades, 0% win rate, -4.60% expectancy) — the open positions look like a continuation of an already-negative pattern for this name, though with n=2 closed that's a hypothesis, not a result.
- Several tickers have **multiple simultaneous tranches at the same unrealized P&L%** (e.g., AZN ×6 at -2.73%, EBAY ×4 at -2.00%, ITUB ×2 at -2.97%) — concentration risk in a handful of names, not diversification, even though position *count* looks high.
- This is situational color only — none of it feeds the point-4 proposal, since none of these trades have a confirmed outcome yet.

## 6. QSR shadow comparison note

Qualitatively notable, observational only:

- **Ticker overlap is low**: legacy shadow candidates (AEP, DDOG, IBKR, MO, AZN, SU, COHR, MAR, TDG, TRP, BTI, ENB — 12 names) and real newlyA entries (CIEN, SU, NOK, AEP, TRP, BTI, EBAY, HON, ITUB, NKE, ENB — 11 names) share only **5 tickers** (AEP, SU, TRP, BTI, ENB). The two selection methods are picking largely different names, not just re-timing the same ones.
- **Frequency**: newlyA is producing far more actual entries (12 closed + 70 open = 82) than legacy would have generated as candidates (28 events total) over the comparable window — newlyA fires much more often.
- Interesting aside: legacy shadow would have bought **MO**, which is one of the worst-performing symbols in the closed-trade history (-5.97% expectancy, 0% win rate, n=8) — newlyA did not select it. Suggestive that newlyA may be screening out at least one historically bad name, but this is a single-symbol anecdote, not a scored result.
- The shadow window is **still running (ends 2026-08-23)** — it hasn't closed yet, so a full outcome-based backtest comparison can't be requested until then. None of this is evidence for the point-4 proposal, per the instructions.