Walk-Forward Parameter Validation

CMT Ch 39 -- In-Sample / Out-of-Sample Split | FL2 VWAP Band Strategy
MODERATE: Some stability. Best IS combo (LB1_UB4) ranks #4 OOS. Spearman=0.902. Use with caution.
0.902
Spearman Rank Correlation
149.7%
CV In-Sample
43.8%
CV Out-of-Sample
0.232
Raw Sharpe (Best Combo)
0
Deflated Sharpe (Ch 35)
100.0%
Multiple Testing Haircut

Data Split

WindowDatesEntriesBest ComboBest P&L
In-Sample (IS) 2026-07-06 to 2026-07-16 115 LB1 / UB4 $+89.68
Out-of-Sample (OOS) 2026-07-17 to 2026-07-27 24 LB1 / UB5 $+98.27
IS best combo (LB1 / UB4) ranked #4 out of 12 in OOS. Marginal holdover -- needs more data.

Combo-by-Combo Comparison (IS vs OOS)

Combo IS P&LIS Win%IS Rank OOS P&LOOS Win%OOS Rank Rank Shift
LB1 / UB2 $+2.12 42.6% #9 $+27.68 37.5% #10 -1
LB1 / UB3 $+53.65 42.6% #5 $+54.28 37.5% #7 -2
LB1 / UB4 $+89.68 42.6% #1 $+81.65 37.5% #4 -3
LB1 / UB5 $+87.88 41.7% #2 $+98.27 37.5% #1 +1
LB2 / UB2 $-30.42 55.7% #11 $+25.56 37.5% #11 =
LB2 / UB3 $+26.67 54.8% #8 $+52.16 37.5% #8 =
LB2 / UB4 $+66.21 53.0% #4 $+79.53 37.5% #5 -1
LB2 / UB5 $+68.35 52.2% #3 $+96.15 37.5% #2 +1
LB3 / UB2 $-65.29 72.2% #12 $+25.55 54.2% #12 =
LB3 / UB3 $-0.10 68.7% #10 $+50.04 41.7% #9 +1
LB3 / UB4 $+38.04 67.0% #7 $+77.69 41.7% #6 +1
LB3 / UB5 $+44.76 66.1% #6 $+93.03 37.5% #3 +3

How to Read This (CMT Ch 35 + 39)

Spearman Rank Correlation: Measures if combos that rank well in IS also rank well in OOS. +1.0 = perfect agreement, 0.0 = random, -1.0 = completely inverted. Above 0.5 = robust.

Coefficient of Variation (CV): stdev/|mean| of P&L across combos. Below 50% = stable parameter landscape. Above 100% = fragile, small changes in parameters flip sign of P&L.

Deflated Sharpe Ratio: Corrects the best combo's Sharpe for the fact that you tested 12 combos. Expected noise Sharpe for 12 tests = 2.229. Your best combo must beat this threshold to have statistical significance. Haircut of 100.0% means 100.0% of apparent performance may be luck.

Parameter Stability: If the best IS combo ranks in the top 3 OOS AND Spearman > 0.5, you likely have a real edge. If it falls to bottom half, you're curve-fitting.