Results14 Oct 2026ema_momentum_results.md

EMA Momentum (ROC), Acceleration & Spread Dynamics — Binance 1h Perps

Executive answer

EMA momentum does contain timing information: a momentum-Reversal signal fires a median of 9h before the bullish crossover and 9h before the bearish one, and 79% / 80% of them are followed by the matching crossover within 24h — so the lead time is real and roughly symmetric between directions. It is not economically decisive: the volatility-scaled false-turn rate is ≈ 50% / 48% (about a coin flip), the directional forward-return edge is small, and delaying entry improves the result rather than hurting it.

On the pre-registered ema12_sma20 long-short book at 6 bps/side, the combined early-warning strategy returned -46.91% OOS (Sharpe -2.16, maxDD -53.63%, 6458 trades) versus the conventional crossover baseline 16.09% (Sharpe 0.67, maxDD -35.61%) and buy & hold OOS 42.29% (Sharpe 1.12).

Walk-forward selection (choose the configuration on each train window, test on the next unseen window) gives the best strategy confirmed a Sharpe of -0.42 with only 50% positive folds. Only 26 of 108 (config, feature) 24h rank-IC cells reach |IC| > 0.02, and the incremental regression over a baseline that already holds the spread and crossover state adds little R².

Verdict. Earlier signals exist, are measurable, and are directionally informative — but the incremental predictive information is small, the false-turn rate is high, and the economic edge does not survive realistic costs and walk-forward selection. The hypotheses are tested honestly below; the negative result is reported rather than over-fit away.

1. Environment & data

Reuses quant_lab.analytics.ema_backtest (loader, execution, costs, funding) and the causal regime classifier. New code: quant_lab.analytics.ema_momentum (features, early-warning events, lead-time/false-turn analysis, strategy targets) and scripts/ema_momentum_study.py. No native 1h bars — 1h is resampled from 1m (open=first, high=max, low=min, close=last, bucket labelled by open time). ATR = Wilder RMA of true range (period 14), initialized with min_periods=period so the first 14 bars are NaN.

symbol bars start end dup ts gaps max gap (h) bad OHLC ATR warm-up funding events
BTCUSDT 15,312 2025-01-01 2026-09-30 0 0 1 0 13 1,914
ETHUSDT 15,312 2025-01-01 2026-09-30 0 0 1 0 13 1,914
SOLUSDT 15,312 2025-01-01 2026-09-30 0 0 1 0 13 1,914
ZECUSDT 15,312 2025-01-01 2026-09-30 0 0 1 0 13 1,914
NEARUSDT 15,312 2025-01-01 2026-09-30 0 0 1 0 13 1,914
XRPUSDT 15,312 2025-01-01 2026-09-30 0 0 1 0 13 1,914
OGNUSDT 15,312 2025-01-01 2026-09-30 0 0 1 0 13 2,567
HYPEUSDT 11,725 2025-05-30 2026-09-30 0 0 1 0 13 2,930
BNBUSDT 15,312 2025-01-01 2026-09-30 0 0 1 0 13 765
DOGEUSDT 15,312 2025-01-01 2026-09-30 0 0 1 0 13 1,914

Funding timestamps carry sub-minute jitter and are truncated to the hour before joining. HYPEUSDT listings start later than the others. Feature values are NaN during warm-up and excluded from every statistic; no rows are dropped or repaired.

2. Experiment A — predictive diagnostics (before trading)

Rank IC of each feature against forward returns, computed per symbol and then equal-weighted (this avoids cross-sectional contamination from pooling assets at different price levels). |IC| > 0.02 is marked nice-to-have; only 26 of 108 24h cells clear it. IC is a dependence measure, not evidence of a tradable edge.

24h rank IC by configuration and feature

config roc_1 roc_12 accel_1 spread svel_1 svel_12
ema10_ema20 -0.012 -0.019 -0.004 -0.019 0.003 0.010
ema10_ema30 -0.012 -0.019 -0.004 -0.027 0.006 0.012
ema10_sma20 -0.012 -0.019 -0.004 -0.009 0.002 0.004
ema12_ema26 -0.014 -0.021 -0.004 -0.025 0.006 0.013
ema12_ema30 -0.014 -0.021 -0.004 -0.028 0.007 0.013
ema12_sma20 -0.014 -0.021 -0.004 -0.008 0.001 0.003
ema15_ema30 -0.017 -0.023 -0.004 -0.030 0.009 0.015
ema15_sma20 -0.017 -0.023 -0.004 -0.005 0.001 0.002
ema20_ema50 -0.022 -0.027 -0.003 -0.043 0.009 0.013
ema20_sma20 -0.022 -0.027 -0.003 0.010 -0.001 -0.003
ema25_sma20 -0.027 -0.031 -0.003 0.026 -0.004 -0.009
ema30_sma20 -0.031 -0.034 -0.003 0.032 -0.006 -0.012
ema40_sma20 -0.038 -0.040 -0.002 0.037 -0.008 -0.011
ema50_sma20 -0.044 -0.045 -0.002 0.042 -0.007 -0.010
ema5_ema20 -0.008 -0.016 -0.005 -0.016 0.001 0.006
ema5_sma20 -0.008 -0.016 -0.005 -0.010 0.001 0.003
ema8_ema20 -0.011 -0.018 -0.004 -0.018 0.002 0.008
ema8_sma20 -0.011 -0.018 -0.004 -0.009 0.001 0.004

Bootstrap 95% CI on the primary config (per-symbol IC then averaged)

feature horizon rank IC 95% CI
roc_1 1h -0.033 [-0.038, -0.029]
roc_1 24h -0.014 [-0.029, -0.000]
accel_1 1h -0.024 [-0.029, -0.019]
accel_1 24h -0.004 [-0.008, -0.000]
spread 1h -0.014 [-0.019, -0.010]
spread 24h -0.008 [-0.024, 0.006]
svel_1 1h -0.023 [-0.027, -0.017]
svel_1 24h 0.001 [-0.008, 0.010]

Conditional forward returns (primary config, pooled)

feature condition horizon n hit mean median
roc_1 ROC>0 1h 74,533 47.5% 0.008% -0.013%
roc_1 ROC<0 1h 74,870 51.0% 0.001% 0.017%
roc_1 ROC>0 6h 74,518 48.6% 0.060% -0.026%
roc_1 ROC<0 6h 74,835 50.7% -0.009% 0.024%
roc_1 ROC>0 24h 74,415 48.4% 0.174% -0.087%
roc_1 ROC<0 24h 74,758 50.4% 0.027% 0.036%
roc_1 ROC>0 72h 74,253 47.6% 0.303% -0.292%
roc_1 ROC<0 72h 74,440 49.7% 0.266% -0.013%
spread spread>0 1h 74,748 48.2% 0.012% -0.003%
spread spread<0 1h 74,585 50.3% -0.003% 0.006%
spread spread>0 6h 74,708 48.6% 0.065% -0.026%
spread spread<0 6h 74,575 50.6% -0.015% 0.024%
spread spread>0 24h 74,610 48.3% 0.167% -0.093%
spread spread<0 24h 74,493 50.5% 0.029% 0.040%
spread spread>0 72h 74,443 47.5% 0.307% -0.298%
spread spread<0 72h 74,180 49.7% 0.253% -0.013%
svel_1 svel>0 1h 74,762 48.0% 0.004% -0.003%
svel_1 svel<0 1h 74,561 50.4% 0.005% 0.008%
svel_1 svel>0 6h 74,756 49.8% 0.040% 0.000%
svel_1 svel<0 6h 74,517 49.4% 0.010% 0.000%
svel_1 svel>0 24h 74,651 49.5% 0.115% -0.018%
svel_1 svel<0 24h 74,442 49.3% 0.080% -0.027%
svel_1 svel>0 72h 74,427 48.7% 0.299% -0.152%
svel_1 svel<0 72h 74,186 48.5% 0.260% -0.172%
accel_1 accel>0 1h 74,473 48.1% 0.004% -0.003%
accel_1 accel<0 1h 74,920 50.4% 0.005% 0.008%
accel_1 accel>0 6h 74,448 49.1% 0.022% -0.007%
accel_1 accel<0 6h 74,895 50.1% 0.029% 0.007%
accel_1 accel>0 24h 74,353 49.3% 0.095% -0.029%
accel_1 accel<0 24h 74,810 49.5% 0.105% -0.012%
accel_1 accel>0 72h 74,101 48.8% 0.304% -0.133%
accel_1 accel<0 72h 74,582 48.4% 0.264% -0.178%

Incremental information over a baseline with the current spread + crossover state

OLS of forward returns on {spread, crossover_state} (baseline) vs the same plus the feature(s); HAC (Newey–West) errors with maxlags = horizon. ΔR² is the incremental explanatory power; t_add is the HAC t-statistic of the first added regressor.

horizon added n R² base R² full ΔR² t_add
1h roc_1 149,333 0.0001 0.0004 0.00026 -2.83
1h accel_1 149,333 0.0001 0.0001 0.00000 -0.01
1h svel_1 149,323 0.0001 0.0002 0.00003 -1.90
1h roc_1+accel_1+svel_1 149,323 0.0001 0.0005 0.00034 -2.93
6h roc_1 149,283 0.0007 0.0010 0.00028 -1.79
6h accel_1 149,283 0.0007 0.0007 0.00001 -0.58
6h svel_1 149,273 0.0007 0.0008 0.00004 1.34
6h roc_1+accel_1+svel_1 149,273 0.0007 0.0015 0.00071 -2.17
24h roc_1 149,103 0.0005 0.0011 0.00066 -2.06
24h accel_1 149,103 0.0005 0.0005 0.00001 -1.17
24h svel_1 149,093 0.0005 0.0005 0.00001 0.81
24h roc_1+accel_1+svel_1 149,093 0.0005 0.0017 0.00125 -2.23
72h roc_1 148,623 0.0001 0.0005 0.00045 -1.70
72h accel_1 148,623 0.0001 0.0001 0.00000 -0.71
72h svel_1 148,613 0.0001 0.0001 0.00001 1.06
72h roc_1+accel_1+svel_1 148,613 0.0001 0.0010 0.00089 -1.80

Consistency across symbols, years and regimes (primary config, 24h)

symbol IC roc_1 IC spread
BTCUSDT -0.016 -0.005
ETHUSDT -0.031 -0.030
SOLUSDT -0.004 0.010
ZECUSDT 0.037 0.031
NEARUSDT -0.011 -0.004
XRPUSDT -0.026 -0.024
OGNUSDT -0.047 -0.038
HYPEUSDT -0.011 -0.002
BNBUSDT -0.021 -0.014
DOGEUSDT -0.010 -0.006
year n IC roc_1 IC spread
2025 84,013 -0.025 -0.013
2026 65,520 -0.000 -0.001
regime n IC spread (24h, n-weighted)
range_lowvol 52,837 0.023
range_highvol 48,620 -0.050
up_lowvol 11,607 0.045
down_lowvol 11,588 0.053
down_highvol 11,449 -0.047
up_highvol 11,402 0.026

Feature cross-correlation (primary config, finite overlap): roc_1/spread +0.64, roc_1/svel_1 +0.51, accel_1/roc_1 +0.28, accel_1/spread -0.10, accel_1/svel_1 +0.34, spread/svel_1 +0.09.

3. Experiments B & C — early-warning signals, lead time & false turns

Conditions (bullish; bearish mirrored):

The reference event is the conventional crossover of the same moving averages. median_lead_h is the median hours from the signal to that crossover; false_turn_1 is the share where price moves 1×ATR against the signal before moving 1×ATR in favour within 24h. Because each signal sits on the opposite side of the crossover it predicts, the next crossover of those two MAs is, by construction, always the matching direction (wrong_first ≈ 0); the real failure modes are distance (how far the cross is) and reversal (price moving against the signal first), which is why both are reported.

Lead-time summary — primary config ema12_sma20

signal n median lead (h) within 6h within 12h within 24h within 48h hit 24h mean 24h false-turn 1σ MFE/MAE (ATR)
decel_bull 233,455 15 19% 42% 69% 88% 50.9% 0.035% 50% 2.60/2.83
decel_bear 232,633 15 19% 42% 69% 88% 51.4% -0.205% 49% 2.61/3.04
reversal_bull 127,882 9 41% 60% 79% 93% 49.2% -0.043% 50% 2.73/2.82
reversal_bear 125,653 9 42% 61% 80% 93% 51.5% -0.011% 48% 2.73/2.78
converge_bull 168,953 13 27% 48% 71% 89% 50.0% 0.044% 51% 2.73/2.81
converge_bear 167,680 13 27% 48% 70% 88% 51.2% -0.117% 49% 2.71/2.91
combined_bull 116,908 8 42% 62% 79% 93% 49.4% -0.032% 50% 2.72/2.82
combined_bear 114,687 8 43% 62% 80% 93% 51.8% -0.011% 48% 2.71/2.78
confirmed_bull 76,551 9 40% 60% 79% 92% 49.9% 0.005% 50% 2.73/2.79
confirmed_bear 75,777 9 41% 60% 79% 92% 51.7% -0.037% 48% 2.69/2.79

Across all configurations (median lead / within-24h / false-turn 1σ)

config signal n median lead within 24h false-turn 1σ mean 24h
ema5_sma20 decel_bull 14,186 13 80% 51% 0.017%
ema5_sma20 decel_bear 14,431 13 80% 49% -0.202%
ema5_sma20 reversal_bull 11,528 7 88% 50% -0.036%
ema5_sma20 reversal_bear 11,427 7 88% 49% -0.128%
ema5_sma20 converge_bull 11,538 9 85% 51% -0.014%
ema5_sma20 converge_bear 11,583 9 85% 49% -0.153%
ema5_sma20 combined_bull 10,825 7 88% 50% -0.032%
ema5_sma20 combined_bear 10,776 7 87% 49% -0.148%
ema5_sma20 confirmed_bull 6,666 7 88% 50% -0.042%
ema5_sma20 confirmed_bear 6,653 7 87% 49% -0.145%
ema8_sma20 decel_bull 15,189 14 78% 51% 0.038%
ema8_sma20 decel_bear 15,369 14 78% 49% -0.221%
ema8_sma20 reversal_bull 9,396 6 89% 50% -0.061%
ema8_sma20 reversal_bear 9,405 6 88% 48% -0.092%
ema8_sma20 converge_bull 9,986 10 84% 51% 0.009%
ema8_sma20 converge_bear 10,099 9 84% 49% -0.127%
ema8_sma20 combined_bull 8,636 6 89% 50% -0.068%
ema8_sma20 combined_bear 8,630 6 88% 48% -0.098%
ema8_sma20 confirmed_bull 5,203 6 89% 50% -0.038%
ema8_sma20 confirmed_bear 5,297 6 88% 49% -0.103%
ema10_sma20 decel_bull 15,650 14 79% 50% 0.023%
ema10_sma20 decel_bear 15,898 13 78% 49% -0.224%
ema10_sma20 reversal_bull 8,533 5 90% 51% -0.082%
ema10_sma20 reversal_bear 8,481 5 89% 48% -0.042%
ema10_sma20 converge_bull 9,545 9 85% 51% 0.018%
ema10_sma20 converge_bear 9,637 9 84% 48% -0.116%
ema10_sma20 combined_bull 7,676 5 90% 51% -0.056%
ema10_sma20 combined_bear 7,623 5 89% 48% -0.041%
ema10_sma20 confirmed_bull 4,766 6 90% 50% -0.023%
ema10_sma20 confirmed_bear 4,801 6 89% 48% -0.081%
ema12_sma20 decel_bull 16,012 13 81% 50% 0.034%
ema12_sma20 decel_bear 16,191 13 80% 49% -0.220%
ema12_sma20 reversal_bull 7,737 5 91% 51% -0.106%
ema12_sma20 reversal_bear 7,798 5 91% 48% -0.034%
ema12_sma20 converge_bull 9,350 9 86% 51% -0.005%
ema12_sma20 converge_bear 9,411 9 85% 49% -0.107%
ema12_sma20 combined_bull 6,765 5 92% 51% -0.090%
ema12_sma20 combined_bear 6,807 5 91% 48% -0.053%
ema12_sma20 confirmed_bull 4,403 5 92% 50% -0.027%
ema12_sma20 confirmed_bear 4,452 5 91% 49% -0.063%
ema15_sma20 decel_bull 16,152 12 86% 50% 0.034%
ema15_sma20 decel_bear 16,306 12 85% 49% -0.219%
ema15_sma20 reversal_bull 6,737 4 93% 50% -0.063%
ema15_sma20 reversal_bear 6,864 4 93% 48% 0.006%
ema15_sma20 converge_bull 9,216 9 89% 51% 0.019%
ema15_sma20 converge_bear 9,347 8 89% 48% -0.068%
ema15_sma20 combined_bull 5,474 4 94% 50% -0.040%
ema15_sma20 combined_bear 5,582 4 94% 48% -0.016%
ema15_sma20 confirmed_bull 3,804 4 95% 50% -0.030%
ema15_sma20 confirmed_bear 3,914 4 95% 48% -0.023%
ema20_sma20 decel_bull 14,050 9 96% 50% -0.011%
ema20_sma20 decel_bear 14,029 9 95% 50% -0.166%
ema20_sma20 reversal_bull 5,110 6 92% 51% -0.099%
ema20_sma20 reversal_bear 5,269 6 93% 48% -0.055%
ema20_sma20 converge_bull 9,186 8 94% 51% -0.009%
ema20_sma20 converge_bear 9,294 8 94% 49% -0.174%
ema20_sma20 combined_bull 3,505 4 95% 51% -0.104%
ema20_sma20 combined_bear 3,631 4 96% 48% -0.034%
ema20_sma20 confirmed_bull 2,728 4 95% 52% -0.115%
ema20_sma20 confirmed_bear 2,820 4 95% 49% -0.082%
ema25_sma20 decel_bull 10,229 8 98% 51% -0.119%
ema25_sma20 decel_bear 10,291 8 98% 50% -0.160%
ema25_sma20 reversal_bull 4,129 10 90% 52% -0.134%
ema25_sma20 reversal_bear 4,239 10 89% 49% -0.047%
ema25_sma20 converge_bull 8,553 10 82% 51% -0.069%
ema25_sma20 converge_bear 8,682 10 81% 50% -0.129%
ema25_sma20 combined_bull 2,617 7 94% 52% -0.108%
ema25_sma20 combined_bear 2,690 7 93% 48% -0.064%
ema25_sma20 confirmed_bull 2,242 8 91% 52% -0.148%
ema25_sma20 confirmed_bear 2,312 8 91% 50% -0.157%
ema30_sma20 decel_bull 7,589 8 96% 51% -0.085%
ema30_sma20 decel_bear 7,548 8 96% 49% -0.113%
ema30_sma20 reversal_bull 3,579 13 79% 52% -0.195%
ema30_sma20 reversal_bear 3,701 12 80% 49% 0.012%
ema30_sma20 converge_bull 7,871 16 64% 52% 0.013%
ema30_sma20 converge_bear 7,908 16 64% 51% -0.128%
ema30_sma20 combined_bull 2,237 9 84% 51% -0.209%
ema30_sma20 combined_bear 2,345 9 84% 49% 0.029%
ema30_sma20 confirmed_bull 2,015 12 80% 53% -0.215%
ema30_sma20 confirmed_bear 2,073 12 82% 51% -0.063%
ema40_sma20 decel_bull 5,412 8 91% 51% -0.035%
ema40_sma20 decel_bear 5,393 8 90% 49% -0.020%
ema40_sma20 reversal_bull 2,961 15 65% 51% -0.192%
ema40_sma20 reversal_bear 3,103 15 63% 48% -0.064%
ema40_sma20 converge_bull 7,368 26 48% 51% 0.072%
ema40_sma20 converge_bear 7,445 27 47% 51% -0.187%
ema40_sma20 combined_bull 1,908 13 68% 51% -0.164%
ema40_sma20 combined_bear 2,078 13 68% 49% -0.033%
ema40_sma20 confirmed_bull 1,708 15 65% 51% -0.153%
ema40_sma20 confirmed_bear 1,775 15 64% 49% 0.014%
ema50_sma20 decel_bull 4,479 8 89% 51% -0.026%
ema50_sma20 decel_bear 4,525 7 87% 49% -0.081%
ema50_sma20 reversal_bull 2,626 15 62% 51% -0.138%
ema50_sma20 reversal_bear 2,673 15 59% 47% -0.134%
ema50_sma20 converge_bull 7,186 31 42% 51% 0.091%
ema50_sma20 converge_bear 7,472 32 39% 51% -0.191%
ema50_sma20 combined_bull 1,769 14 63% 50% -0.119%
ema50_sma20 combined_bear 1,864 14 61% 48% -0.167%
ema50_sma20 confirmed_bull 1,513 16 61% 50% -0.064%
ema50_sma20 confirmed_bear 1,585 15 59% 47% -0.058%
ema5_ema20 decel_bull 14,026 15 71% 51% 0.026%
ema5_ema20 decel_bear 14,159 15 71% 49% -0.207%
ema5_ema20 reversal_bull 11,469 9 81% 50% -0.038%
ema5_ema20 reversal_bear 11,155 8 81% 49% -0.095%
ema5_ema20 converge_bull 12,544 11 78% 51% -0.011%
ema5_ema20 converge_bear 12,305 11 77% 49% -0.165%
ema5_ema20 combined_bull 11,467 9 81% 50% -0.038%
ema5_ema20 combined_bear 11,155 8 81% 49% -0.095%
ema5_ema20 confirmed_bull 7,489 9 80% 50% -0.002%
ema5_ema20 confirmed_bear 7,280 9 81% 49% -0.088%
ema8_ema20 decel_bull 14,598 18 63% 51% 0.045%
ema8_ema20 decel_bear 14,568 18 63% 49% -0.219%
ema8_ema20 reversal_bull 9,312 9 78% 50% -0.020%
ema8_ema20 reversal_bear 9,075 9 78% 48% -0.026%
ema8_ema20 converge_bull 10,889 13 72% 51% 0.035%
ema8_ema20 converge_bear 10,656 13 71% 48% -0.123%
ema8_ema20 combined_bull 9,311 9 78% 50% -0.020%
ema8_ema20 combined_bear 9,075 9 78% 48% -0.026%
ema8_ema20 confirmed_bull 5,928 10 78% 50% 0.028%
ema8_ema20 confirmed_bear 5,823 10 77% 49% -0.061%
ema10_ema20 decel_bull 14,832 20 60% 50% 0.046%
ema10_ema20 decel_bear 14,835 20 59% 49% -0.224%
ema10_ema20 reversal_bull 8,402 9 77% 50% -0.028%
ema10_ema20 reversal_bear 8,132 9 78% 47% 0.025%
ema10_ema20 converge_bull 10,104 14 69% 50% 0.063%
ema10_ema20 converge_bear 9,945 14 68% 48% -0.136%
ema10_ema20 combined_bull 8,401 9 77% 50% -0.028%
ema10_ema20 combined_bear 8,132 9 78% 47% 0.025%
ema10_ema20 confirmed_bull 5,261 10 76% 50% 0.053%
ema10_ema20 confirmed_bear 5,174 10 76% 47% 0.003%
ema10_ema30 decel_bull 13,998 25 49% 50% 0.069%
ema10_ema30 decel_bear 13,783 25 49% 49% -0.244%
ema10_ema30 reversal_bull 8,324 13 68% 50% 0.009%
ema10_ema30 reversal_bear 7,924 13 69% 47% 0.060%
ema10_ema30 converge_bull 9,712 18 60% 50% 0.116%
ema10_ema30 converge_bear 9,381 18 60% 48% -0.052%
ema10_ema30 combined_bull 8,324 13 68% 50% 0.009%
ema10_ema30 combined_bear 7,924 13 69% 47% 0.060%
ema10_ema30 confirmed_bull 5,216 14 68% 49% 0.103%
ema10_ema30 confirmed_bear 5,027 14 68% 47% -0.005%
ema12_ema26 decel_bull 14,470 25 50% 50% 0.077%
ema12_ema26 decel_bear 14,279 25 50% 49% -0.232%
ema12_ema26 reversal_bull 7,653 12 70% 50% -0.019%
ema12_ema26 reversal_bear 7,320 12 71% 47% 0.074%
ema12_ema26 converge_bull 9,321 18 61% 50% 0.113%
ema12_ema26 converge_bear 9,082 18 60% 48% -0.060%
ema12_ema26 combined_bull 7,652 12 70% 50% -0.019%
ema12_ema26 combined_bear 7,320 12 71% 47% 0.074%
ema12_ema26 confirmed_bull 4,779 12 70% 49% 0.069%
ema12_ema26 confirmed_bear 4,615 13 69% 47% 0.013%
ema12_ema30 decel_bull 14,180 27 46% 50% 0.102%
ema12_ema30 decel_bear 13,892 27 47% 49% -0.227%
ema12_ema30 reversal_bull 7,604 13 67% 50% -0.009%
ema12_ema30 reversal_bear 7,265 13 68% 47% 0.079%
ema12_ema30 converge_bull 9,230 20 57% 50% 0.118%
ema12_ema30 converge_bear 8,984 20 57% 48% -0.058%
ema12_ema30 combined_bull 7,604 13 67% 50% -0.009%
ema12_ema30 combined_bear 7,265 13 68% 47% 0.079%
ema12_ema30 confirmed_bull 4,755 14 67% 49% 0.059%
ema12_ema30 confirmed_bear 4,589 14 67% 47% 0.021%
ema15_ema30 decel_bull 14,478 29 42% 50% 0.097%
ema15_ema30 decel_bear 14,092 29 43% 49% -0.238%
ema15_ema30 reversal_bull 6,857 14 66% 50% 0.050%
ema15_ema30 reversal_bear 6,461 13 68% 47% 0.125%
ema15_ema30 converge_bull 8,799 22 54% 50% 0.127%
ema15_ema30 converge_bear 8,487 22 54% 49% -0.062%
ema15_ema30 combined_bull 6,857 14 66% 50% 0.050%
ema15_ema30 combined_bear 6,461 13 68% 47% 0.125%
ema15_ema30 confirmed_bull 4,260 14 65% 49% 0.133%
ema15_ema30 confirmed_bear 4,103 15 66% 47% 0.031%
ema20_ema50 decel_bull 13,925 46 26% 49% 0.113%
ema20_ema50 decel_bear 13,044 43 28% 49% -0.216%
ema20_ema50 reversal_bull 5,925 23 52% 49% 0.123%
ema20_ema50 reversal_bear 5,361 20 56% 46% 0.228%
ema20_ema50 converge_bull 8,555 36 37% 49% 0.157%
ema20_ema50 converge_bear 7,962 34 38% 47% -0.073%
ema20_ema50 combined_bull 5,880 23 52% 49% 0.124%
ema20_ema50 combined_bear 5,329 20 56% 46% 0.229%
ema20_ema50 confirmed_bull 3,815 24 50% 48% 0.124%
ema20_ema50 confirmed_bear 3,484 21 54% 46% 0.275%

False-turn sensitivity to the volatility-scaled threshold (primary config)

signal n 0.5σ 1.0σ 1.5σ
combined_bear 6,807 45% 48% 47%
combined_bull 6,765 48% 51% 48%
confirmed_bear 4,452 47% 49% 47%
confirmed_bull 4,403 48% 50% 48%
converge_bear 9,411 46% 49% 48%
converge_bull 9,350 48% 51% 48%
decel_bear 16,191 45% 49% 49%
decel_bull 16,012 46% 50% 49%
reversal_bear 7,798 46% 48% 47%
reversal_bull 7,737 48% 51% 49%

4. Experiment D — tradable strategy variants

All strategies are latch-based: a bull event opens/holds the long side, a bear event the short side, reversing on the opposite event. Costs and funding use the shared engine; repeated same-direction signals maintain the position. The primary-config comparison below is at 6 bps/side.

strategy side IS ret IS Sharpe OOS ret OOS Sharpe OOS maxDD trades win PF avg trade hold h break-even (bps)
combined both -36.55% -0.77 -46.91% -2.16 -53.63% 6458 45% 0.88 -0.150% 23.1 ≤0
combined long -18.21% -0.28 -8.89% -0.18 -31.26% 6450 45% 0.96 -0.053% 23.0 2.7
combined short -27.90% -0.55 -43.77% -2.25 -51.08% 6456 45% 0.82 -0.246% 23.2 ≤0
confirmed both -34.10% -0.64 -43.09% -1.79 -51.30% 4768 51% 0.87 -0.195% 31.3 ≤0
confirmed long -16.97% -0.22 -5.91% -0.03 -31.26% 4759 50% 0.96 -0.062% 31.2 3.0
confirmed short -25.94% -0.50 -41.47% -2.17 -47.35% 4767 52% 0.79 -0.327% 31.4 ≤0
cross both -26.45% -0.25 16.09% 0.67 -35.61% 8660 33% 1.02 0.020% 17.2 8.1
cross long -7.61% 0.02 36.78% 1.38 -23.01% 8659 32% 1.08 0.096% 17.2 14.3
cross short -23.72% -0.30 -17.00% -0.47 -33.93% 8651 34% 0.95 -0.055% 17.2 1.2
roc both -48.91% -1.22 -55.49% -2.94 -59.27% 7403 42% 0.85 -0.182% 20.1 ≤0
roc long -26.51% -0.54 -16.48% -0.54 -36.32% 7395 42% 0.91 -0.098% 20.1 0.4
roc short -35.30% -0.79 -48.67% -2.59 -52.82% 7401 43% 0.79 -0.267% 20.2 ≤0
velocity both -42.97% -0.75 -72.04% -3.74 -73.25% 7519 51% 0.82 -0.229% 19.8 ≤0
velocity long -21.68% -0.30 -33.60% -1.27 -43.38% 7510 50% 0.88 -0.145% 19.8 ≤0
velocity short -30.77% -0.63 -58.95% -3.37 -61.72% 7518 51% 0.77 -0.312% 19.9 ≤0

Best OOS Sharpe across the configuration grid, per strategy

strategy side best config OOS return OOS Sharpe OOS maxDD configs tested
cross long ema10_ema20 41.41% 1.54 -21.18% 18
cross short ema10_ema20 -14.48% -0.36 -37.74% 18
cross both ema10_ema20 23.78% 0.86 -28.57% 18
roc long ema20_ema50 54.85% 1.78 -26.46% 18
roc short ema20_ema50 -4.13% -0.02 -31.01% 18
roc both ema20_ema50 53.29% 1.67 -27.08% 18
velocity long ema20_ema50 38.72% 1.33 -32.46% 18
velocity short ema20_ema50 -13.90% -0.44 -31.69% 18
velocity both ema20_ema50 22.38% 0.84 -29.96% 18
combined long ema20_ema50 54.56% 1.78 -26.46% 18
combined short ema20_ema50 -4.30% -0.02 -30.98% 18
combined both ema20_ema50 52.74% 1.66 -27.08% 18
confirmed long ema20_ema50 47.05% 1.57 -28.85% 18
confirmed short ema20_ema50 -8.85% -0.22 -33.39% 18
confirmed both ema20_ema50 38.09% 1.27 -28.94% 18

Risk-management variants (primary config, combined, OOS)

risk rule side OOS return OOS Sharpe OOS maxDD trades exposure
none long -8.89% -0.18 -31.26% 6450 50%
none short -43.77% -2.25 -51.08% 6456 50%
none both -46.91% -2.16 -53.63% 6458 100%
minhold12 long -6.18% -0.07 -28.53% 6205 50%
minhold12 short -42.18% -2.11 -51.60% 6211 50%
minhold12 both -43.82% -1.93 -51.75% 6213 100%
atrstop2 long 46.45% 1.24 -35.54% 484 94%
atrstop2 short -46.04% -1.21 -61.73% 280 96%
atrstop2 both -42.06% -2.78 -43.60% 8257 60%

Entry-delay sensitivity (primary combined, long-short)

delay (bars) OOS return OOS Sharpe full-sample return
0 -46.91% -2.16 -66.31%
1 -38.80% -1.63 -53.49%
2 -20.66% -0.69 -51.10%
3 -20.71% -0.68 -54.71%
6 -3.75% 0.04 -57.33%
12 5.85% 0.39 -15.66%

ROC-lookback k sensitivity (primary combined, long-short, OOS)

k OOS return OOS Sharpe trades
1 -46.91% -2.16 6458
3 -4.43% -0.01 4959
6 -3.39% -0.01 2521
12 -1.37% 0.01 73

5. Validation & selection bias

Per fold the configuration with the best train Sharpe is frozen (for each strategy) and evaluated on the next unseen test window; test windows are stitched for the aggregate. Selection is over the full candidate grid, so this is the multiple-testing-honest result.

strategy folds WF OOS ret WF OOS Sharpe WF OOS maxDD mean test Sharpe folds positive
cross 4 -48.62% -1.12 -62.88% -1.19 0%
roc 4 -46.16% -1.41 -47.44% -1.58 25%
velocity 4 -49.75% -1.24 -58.47% -1.30 25%
combined 4 -49.14% -1.54 -51.56% -1.71 25%
confirmed 4 -21.58% -0.42 -41.86% -0.83 50%
strategy fold train window test window chosen config train Sharpe val Sharpe test Sharpe test return
combined 0 2025-01-01→2025-07-01 2025-09-01→2025-12-01 ema20_ema50 0.84 3.18 -4.35 -41.48%
combined 1 2025-04-01→2025-10-01 2025-12-01→2026-03-01 ema20_ema50 1.27 -5.46 2.51 29.51%
combined 2 2025-07-01→2026-01-01 2026-03-01→2026-06-01 ema40_sma20 1.95 -3.63 -2.16 -15.27%
combined 3 2025-10-01→2026-04-01 2026-06-01→2026-09-01 ema50_sma20 0.38 -0.42 -2.82 -20.80%
confirmed 0 2025-01-01→2025-07-01 2025-09-01→2025-12-01 ema20_sma20 1.73 -0.20 0.32 1.14%
confirmed 1 2025-04-01→2025-10-01 2025-12-01→2026-03-01 ema20_ema50 0.97 -3.02 2.00 24.01%
confirmed 2 2025-07-01→2026-01-01 2026-03-01→2026-06-01 ema40_sma20 2.63 -3.72 -1.10 -9.06%
confirmed 3 2025-10-01→2026-04-01 2026-06-01→2026-09-01 ema40_sma20 0.81 -0.06 -4.55 -31.24%
cross 0 2025-01-01→2025-07-01 2025-09-01→2025-12-01 ema10_sma20 0.20 1.47 -0.49 -10.50%
cross 1 2025-04-01→2025-10-01 2025-12-01→2026-03-01 ema50_sma20 0.75 -3.32 -1.49 -20.11%
cross 2 2025-07-01→2026-01-01 2026-03-01→2026-06-01 ema5_ema20 0.01 0.68 -1.75 -16.37%
cross 3 2025-10-01→2026-04-01 2026-06-01→2026-09-01 ema12_sma20 -0.22 -0.69 -1.04 -14.08%
roc 0 2025-01-01→2025-07-01 2025-09-01→2025-12-01 ema20_ema50 0.84 3.18 -4.38 -41.75%
roc 1 2025-04-01→2025-10-01 2025-12-01→2026-03-01 ema20_ema50 1.27 -5.49 2.53 29.82%
roc 2 2025-07-01→2026-01-01 2026-03-01→2026-06-01 ema50_sma20 1.75 -2.93 -2.70 -17.33%
roc 3 2025-10-01→2026-04-01 2026-06-01→2026-09-01 ema50_sma20 0.85 -1.63 -1.79 -13.88%
velocity 0 2025-01-01→2025-07-01 2025-09-01→2025-12-01 ema25_sma20 1.25 -4.98 -1.27 -18.78%
velocity 1 2025-04-01→2025-10-01 2025-12-01→2026-03-01 ema20_ema50 1.46 -4.54 1.56 17.65%
velocity 2 2025-07-01→2026-01-01 2026-03-01→2026-06-01 ema8_ema20 0.23 -2.23 -1.16 -11.15%
velocity 3 2025-10-01→2026-04-01 2026-06-01→2026-09-01 ema8_ema20 0.15 0.16 -4.35 -40.81%

6. Costs & break-even

Primary combined strategy: gross OOS return (0 bps, no funding effect net of the series) -25.95%; break-even costs below. Sensitivity at 0/3/6/12/20 bps/side.

strategy scenario bps/side OOS ret OOS Sharpe OOS maxDD fees funding
roc zero 0 -34.48% -1.46 -42.77% 0.00% -0.08%
roc low 3 -46.00% -2.20 -51.51% 44.39% -0.08%
roc base 6 -55.49% -2.94 -59.27% 88.78% -0.08%
roc high 12 -69.77% -4.42 -71.72% 177.55% -0.08%
roc extreme 20 -81.95% -6.38 -82.61% 295.92% -0.08%
velocity zero 0 -58.61% -2.53 -61.02% 0.00% -1.56%
velocity low 3 -65.98% -3.14 -67.71% 45.08% -1.56%
velocity base 6 -72.04% -3.74 -73.25% 90.17% -1.56%
velocity high 12 -81.11% -4.95 -81.66% 180.34% -1.56%
velocity extreme 20 -88.81% -6.55 -88.93% 300.56% -1.56%
combined zero 0 -25.95% -0.93 -39.41% 0.00% -0.09%
combined low 3 -37.30% -1.54 -46.60% 38.72% -0.09%
combined base 6 -46.91% -2.16 -53.63% 77.44% -0.09%
combined high 12 -61.94% -3.38 -65.39% 154.87% -0.09%
combined extreme 20 -75.59% -5.00 -77.10% 258.12% -0.09%
cross zero 0 83.29% 2.03 -26.68% 0.00% 0.81%
cross low 3 45.88% 1.35 -31.29% 51.93% 0.81%
cross base 6 16.09% 0.67 -35.61% 103.86% 0.81%
cross high 12 -26.48% -0.69 -46.70% 207.72% 0.81%
cross extreme 20 -60.02% -2.50 -65.56% 346.20% 0.81%

Break-even fee+slippage per side (primary config) by strategy/side: cross/long 14.3, cross/short 1.2, cross/both 8.1, roc/long 0.4, roc/short ≤0, roc/both ≤0, velocity/long ≤0, velocity/short ≤0, velocity/both ≤0, combined/long 2.7, combined/short ≤0, combined/both ≤0, confirmed/long 3.0, confirmed/short ≤0, confirmed/both ≤0 bps.

7. Market-regime analysis (primary combined, OOS)

regime bars mean symbol return % positive symbols
down_highvol 4,948 9.10% 60%
up_highvol 5,405 0.65% 50%
up_lowvol 4,721 -7.42% 40%
down_lowvol 5,701 -13.50% 20%
range_highvol 21,130 -19.11% 10%
range_lowvol 23,615 -28.13% 0%

Regimes from the causal classifier (trend = volatility-normalised trailing 24h move, ±1σ; vol = realised vol vs trailing median), shifted one bar. OOS only.

8. Robustness & statistical analysis

Bootstrap (block=72h, 1000 resamples) on the primary combined OOS returns: Sharpe mean -2.18, 95% CI [-4.33, 0.05]; total-return 95% CI [-70.66%, -0.63%].

Profit concentration (primary combined, pooled trades): 6458 trades; top 5% = 47% of gross profit; best 84.49%, worst -51.92%.

Per-symbol OOS (primary combined)

symbol OOS return OOS Sharpe OOS maxDD trades win PF
DOGEUSDT -26.62% -0.29 -57.21% 647 45% 0.77
HYPEUSDT -33.35% -0.14 -70.05% 502 46% 0.95
XRPUSDT -45.76% -0.97 -69.76% 662 45% 0.80
SOLUSDT -45.78% -0.99 -54.48% 675 46% 0.78
BTCUSDT -53.33% -2.14 -59.79% 681 45% 0.89
ETHUSDT -55.61% -1.61 -65.19% 673 43% 0.98
NEARUSDT -60.92% -0.74 -78.98% 671 46% 1.02
BNBUSDT -63.26% -2.91 -66.63% 661 45% 0.69
ZECUSDT -67.91% -0.68 -86.55% 623 46% 0.91
OGNUSDT -78.71% -1.65 -82.64% 663 43% 0.95

Baselines over the same OOS window: buy & hold 42.29% (Sharpe 1.12); a simple 24h price-momentum long-short baseline -14.01% (Sharpe -0.22); no-trade 0%.

Multiple testing: the strategy grid is 270 (strategy, side, config) cells, and the diagnostics add 432 (config, feature, horizon) IC cells. The best observed values are order statistics; walk-forward selection on train folds is the honest evaluation and is reported above.

9. Limitations

10. Answers to the study's final questions

  1. Does EMA ROC predict before crossovers? Weakly, but with a genuine and roughly symmetric lead. Reversal signals precede the same-direction cross with a median lead of 9h (bull) / 9h (bear); rank IC vs 24h return is small (-0.019 average over configs) and conditional mean returns are near zero.
  2. Does acceleration add information beyond ROC? Marginally. The incremental R² from adding accel_1 on top of the spread + crossover baseline is 0.00001 at 24h; it does not change the sign of the conclusion.
  3. Does normalized spread velocity add beyond the current spread? Slightly more than acceleration: ΔR² ≈ 0.00001 at 24h, with HAC t-stats that are positive but modest.
  4. Does the combined signal beat its components? Not out-of-sample after costs — see the strategy table; the components are close together and the combined rule trades a different amount rather than dominating.
  5. Lead time and false turns? Median lead ≈ 9h across signals; the 1×ATR false-turn rate is ≈ 49% on average — high — and lead time and reliability trade off sharply (best reliability is the short-lead combined/confirmed bull signal).
  6. Does earlier entry improve expectancy? No — the opposite. In the entry-delay table the un-delayed early signal is the worst (OOS ≈ -47% at 6 bps); adding a 6–12 bar delay lifts it to roughly break-even. The earliest signal is the noisiest, so early entry destroys expectancy rather than creating it, and the crossover baseline is the better of the two.
  7. Do gains survive costs & funding? No strategy clears the base taker cost with a stable sign; break-even costs are modest relative to the 6 bps assumption (the crossover baseline is the only primary-config variant with a positive break-even on the long side).
  8. Do long and short differ? The diagnostic lead time is nearly symmetric (≈9h either way), but the traded result is not: the long side is far less bad than the short side (primary combined OOS long ≈ -9% vs short ≈ -44%). That is consistent with the sample's upward drift rather than a structural asymmetry in the signal.
  9. Stable across symbols, regimes, folds? No. Per-symbol dispersion is large, regime signs differ, and walk-forward folds are mostly negative.
  10. Is there a repeatable edge? No. H1/H3 are weakly supported diagnostically, H2/H4 are marginal, H5 holds (lead time exists) but with unacceptable false turns, and H6/H7/H8 are not supported for trading. The momentum-deceleration family is rejected as a standalone signal; the diagnostic value of ROC and spread velocity may merit a conditional (regime-gated) retest with finer execution data, but the current evidence does not justify strategy development.

Reproduce: python scripts/ema_momentum_study.py (engine: quant_lab.analytics.ema_backtest; features/signals: quant_lab.analytics.ema_momentum). Artifacts: experiments/reports/ema_momentum_v1.

← All reports · source (.md)