Regime study · 2m intraday bars · 20-SMA + 2% gaps

Markov Regime Dashboard

Rows are today's state, columns tomorrow's. Greener = more likely. Small number = how many times it happened.

NVDA

2026-06-04 to 2026-07-24 · 6805 bars
Trend U 41%Trend D 40%Chop 18%
Trend U
7.4days
persist 0.86 vs 0.41 rand
Trend D
7.1days
persist 0.86 vs 0.40 rand
Chop
2.1days
persist 0.53 vs 0.19 rand
Gap days
6/ 6805
catalyst freq

Transition matrix P(next | current)

Gap UGap DTrend UTrend DChop
Gap U0.0000.0001.0030.0000.000
Gap D0.0000.0000.0000.6720.331
Trend U0.0010.0030.8624220.03950.10280
Trend D0.0010.0000.03770.8623540.11311
Chop0.0010.0000.243000.232910.53662
In-sample only. Trust cells with high counts; gap rows are usually thin. Hypotheses, not trade signals.

Backtest · exit-on-chop vs scale-and-trail

NVDA 2m · 6805 bars · 20bps/side · 1.5% trail · entry: 2 up bars + >VWAP + >=EMA9

DESIGN half

You may peek. In-sample results are NOT evidence.
A · exit on chop
trades
114
win rate
10%
expectancy
-0.438%
avg win
+0.55%
avg loss
-0.54%
total
-39.5%
max DD
-40.4%
B · scale + trail
trades
38
win rate
18%
expectancy
-0.285%
avg win
+0.48%
avg loss
-0.46%
total
-10.3%
max DD
-11.0%

HELD-OUT half

The only scores that count — the rule never saw this data.
A · exit on chopNO EDGE (out-of-sample)
trades
92
win rate
11%
expectancy
-0.419%
avg win
+0.73%
avg loss
-0.56%
total
-32.1%
max DD
-33.0%
B · scale + trailNO EDGE (out-of-sample)
trades
26
win rate
23%
expectancy
-0.188%
avg win
+0.67%
avg loss
-0.44%
total
-4.8%
max DD
-6.4%
A equity
B equity
Read the held-out half only. Out-of-sample, both rules lose — but B loses far less than A (expectancy -0.188% vs -0.419%, drawdown -6% vs -33%). Scaling helps, but this naive trend-entry has no tradeable edge. The weak link is the ENTRY, not the exit — try requiring confirmation before entering. The split (design 60% / held-out 40%) is the guardrail against curve-fitting: a rule that only wins on the half it was designed on is a fantasy. Raise the cost or change the entry and rerun — but never tune a parameter to make the design half prettier.