Each line is one bot's running total of closed-trade P&L. The ribbon underneath is the
market regime on that date, from the same classifier the prediction page shows you.
Trending — calm or steady uptrendChoppy / range-boundBearish / high volatility
Why the version comparison is unsafe
Average P&L per trade, split by model version and by the regime at entry. Hatched cells are
combinations that have never happened.
Regime effect, bot by bot
Average P&L per $1,000 trade in each regime.
The numbers
Actual is the price at the moment the resolver ran, and is
the source of truth. Order-fill is what a real stop/limit
order would have filled at, walked from the daily high/low path. Both are shown because a stop
can gap past its level.
How to read this — and how not to
Regime and version are nearly collinear here. The v5 cutover lands within
days of the market turning choppy, so this page cannot fully separate the two causes — and
neither can any other view of this data. The within-regime comparison above is the closest
available, and it is thin.
Regime is assigned at entry, using SPY versus its 200-day average, the
21-day Kaufman efficiency ratio, and VIX — the same _classify_regime behind the
banner on the prediction page. A trade opened in a trend but closed in chop counts as trending.
Cumulative dollars reward volume, not skill. The bots have very different
trade counts, so their lines are not on comparable footing. The per-trade averages are the
fair comparison.
Open trades are excluded entirely. Only closed, resolved trades appear. If
open positions skew toward losers held to expiry, these totals flatter every bot equally.
This is not a trading rule. The regime split is measured on the same data
that suggested it, and only a handful of calm-uptrend days exist in the whole sample. Gating
entries on regime would need forward out-of-sample confirmation first.