# Learning-loop review — 2026-08-26

336 closed trades reviewed, 200 currently open. Proposal only — nothing here is applied automatically (CLAUDE.md §4: the AI review proposes, dan disposes).

## 1. Is overall expectancy holding?

Yes — pooled expectancy is **+1.77% per trade** across 336 closed trades, profit factor 2.30, positive across essentially every strategy except orb-v1. At n=336 the overall number is well past the ~20-trade minimum and is not noise (Wilson lower bound on win rate is 65.1% vs raw 70.2%, a tight interval). However, remember this pools three accounts with different risk configs (paper-main n=277, Paper02qsr n=22, live-1 n=37) — the qualitative "expectancy is holding" conclusion is fine, but any account-specific claim would need to lean on the per-account breakdown separately.

## 2. Which bucket is the biggest drag?

Among buckets with ≥20 trades, **orb-v1** (n=68) is the clearest drag: expectancy ≈ **-0.00%**, profit factor **0.92** (below 1.0 = losing), win rate 45.6% (Wilson lower bound 34.3%), total P&L **-$67**. Every other strategy bucket with ≥20 trades (qsr n=244, vwap-mr-v1 n=23) is net positive. (The `stop_loss` exit-reason bucket looks even worse, -4.36% expectancy/PF 0.02 on n=87, but that's a tautological bucket — it's just "the trades that lost by hitting a stop" — so it's used below as a diagnostic, not named as the primary drag.)

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

This looks like an **entry-quality problem, not an exit/give-back problem**. Evidence:

- Aggregate **MFE of losers is only 0.79%** — low. Per the doc's own rule, an exit problem requires losers that ran up meaningfully before being given back (high MFE). 0.79% is not that pattern, so we can largely rule out "signal was right, we mismanaged the exit."
- Aggregate **MAE of losers is -4.09%**, i.e. losing trades go deeply against the trade before finally stopping out — consistent with a strategy taking too many low-conviction signals that are simply wrong from early on, not signals that "worked" and reversed late.
- orb-v1's own numbers reinforce this: win rate 45.6% (near coin-flip, Wilson lower 34.3%) with a merely-adequate payoff (1.19) and 0.3-day average hold — the breakout confirmation isn't filtering out enough weak/failed breakouts.

Caveat: the report doesn't break MAE/MFE out per-strategy, so this reasoning leans on the aggregate MAE/MFE pattern plus orb-v1's win-rate/payoff shape, not a strategy-specific MAE/MFE table. That's a real limitation, not a certainty.

## 4. One concrete, small parameter change

**Proposal:** raise orb-v1's `volumeConfirmMult` from **2.5 → 3.0**.

Reasoning: this is an entry-confirmation filter (higher relative-volume requirement on the breakout), which is the natural lever for an entry-quality problem (per point 3) rather than an exit lever like `targetRMultiple`. orb-v1's near-coin-flip win rate with only adequate payoff suggests the current 2.5× volume-confirmation bar is letting through breakouts that don't have real conviction behind them. This is a single, small, reversible tweak to the one strategy that's actually dragging (n=68, clears the 20-trade bar), and it should be judged only on orb-v1 trades entered *after* the change, per rule 4.

No other strategy bucket at ≥20 trades shows a comparable problem, so no other parameter is being touched.

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

- **QSR has a large number of positions with no calendar time-stop** (`maxHoldDays: null`), and several have been open a long time while underwater: **HDB** — 4 tranches, 27–35 days held, all -3.63%; **HON** — several tranches ~14.3 days at **-4.13%**; **ENB** — 13–26 days, -0.90%; **ETR** — ~21–26 days, -0.66%. These are just aging, unresolved risk, not yet evidence of anything — but worth watching since qsr has no time-based exit to force a decision.
- Crypto-trend-v1 BTC/USD and ETH/USD have been open **36.2 days** with no calendar time-stop either.
- One notable open winner: `ETHUSD` (external) at **+28.03%** after 22 days — not part of the closed-trade evidence base, just flagged as an outlier still open.

## 6. QSR shadow comparison note

Purely observational (not usable for point 4): the legacy isTriggered+isBuy/buyZonePct shadow method fired far less often (48 would-buy events, 38 buy-tier + 10 near-miss, across 18 tickers) than the live newlyA method over the comparable/following period (85 closed + 169 open, across 34 tickers) — newlyA is both more prolific and casts a much wider net. Ticker overlap is partial (10 of 18 legacy tickers — AEP, SU, COHR, TRP, BTI, ENB, NKE, ITUB, EBAY, VALE — also appear in newlyA's set), meaning the two selection methods are picking meaningfully different, only partly-overlapping universes rather than the same trades timed differently. The shadow window (2026-08-09 → 2026-08-23) has now closed, so a fuller backtest comparing hypothetical shadow-log outcomes to the real newlyA outcomes can be requested if a rigorous entry-method comparison is wanted — but that would be a separate, explicitly-flagged analysis, not folded into this review's parameter proposal.