Research · Validation

Backtest Results

Honest, cost-inclusive results on held-out data the model never trained on. No cherry-picking.

USDCHF — Held out test set, never trained on (1h, horizon 4, submitted 2026-07-10) -- v2 CORRECTED backtest (supersedes entry 5's inflated Sharpe -- overlapping-forward-return bug, see backtest_signal_model_v2.py docstring). 2021-22 independent robustness window. Sharpe 1.98, +22.7% vs buy-hold +4.9%, max DD -4.0%, 2611 episodes, 60.7% win rate. Confirms the edge on a second window.

Outperformed buy & hold

Test bars

12,473

Trades taken

2,611

Strategy return

22.73%

Buy & hold return

4.87%

Sharpe (annualized)

1.98

Max drawdown

-3.95%

Win rate 60.7%Positive Sharpe and beat buy-and-hold -- a real candidate worth further scrutiny, not yet a confirmed edge on its own.
AUDUSD — Held out test set, never trained on (1D, horizon 12, submitted 2026-07-10) -- REJECTED -- failed robustness (see next entry). Promising on 2024-25 alone: Sharpe 1.09, +18.2% vs buy-hold -1.5%, val AUC 0.63. But did not replicate on the independent 2021-22 window, so treat this as a lucky window, not a real edge.

Outperformed buy & hold

Test bars

508

Trades taken

124

Strategy return

18.17%

Buy & hold return

-1.50%

Sharpe (annualized)

1.09

Max drawdown

-7.67%

Win rate 53.2%Positive Sharpe and beat buy-and-hold -- a real candidate worth further scrutiny, not yet a confirmed edge on its own.
USDCHF — Held out test set, never trained on (1h, horizon 4, submitted 2026-07-10) -- v2 CORRECTED backtest (supersedes entry 4's inflated Sharpe 1.29 -- see backtest_signal_model_v2.py docstring for the overlapping-forward-return bug). 2024-25 held-out window. Still the one real candidate: Sharpe 0.83, +8.2% vs buy-hold -5.5%, max DD -5.2%, 2089 episodes, 56.6% win rate.

Outperformed buy & hold

Test bars

12,452

Trades taken

2,089

Strategy return

8.17%

Buy & hold return

-5.48%

Sharpe (annualized)

0.83

Max drawdown

-5.20%

Win rate 56.6%Positive Sharpe and beat buy-and-hold -- a real candidate worth further scrutiny, not yet a confirmed edge on its own.
USDCHF — Held out test set, never trained on (4h, horizon 24, submitted 2026-07-10) -- REJECTED after v2 correction. Old v1 backtest showed Sharpe 1.39 / +118%, but that was entirely the overlapping-forward-return bug (val AUC was 0.44 -- worse than random). Corrected 2024-25: Sharpe -0.13, -2.2% vs buy-hold -6.4%. No edge.

Outperformed buy & hold

Test bars

3,090

Trades taken

654

Strategy return

-2.17%

Buy & hold return

-6.43%

Sharpe (annualized)

-0.13

Max drawdown

-12.42%

Win rate 48.5%Negative risk-adjusted return, consistent with the win rate above. Confirms no exploitable edge for this pair/horizon after realistic spread costs.
USDCHF — Held out test set, never trained on (4h, horizon 24, submitted 2026-07-10) -- REJECTED after v2 correction, 2021-22 robustness window. Old v1 backtest showed inflated Sharpe from the overlapping-forward-return bug. Corrected: Sharpe -0.42, -6.0% vs buy-hold +4.5%. Confirms no edge on a second window.

Underperformed buy & hold

Test bars

3,118

Trades taken

571

Strategy return

-5.97%

Buy & hold return

4.52%

Sharpe (annualized)

-0.42

Max drawdown

-7.89%

Win rate 52.9%Negative risk-adjusted return, consistent with the win rate above. Confirms no exploitable edge for this pair/horizon after realistic spread costs.
AUDUSD — Held out test set, never trained on (1D, horizon 12, submitted 2026-07-10) -- REJECTED -- robustness failure confirmed. 2021-22 independent window: Sharpe -0.46, -10.5% vs buy-hold -11.6%, val AUC 0.46 (worse than random). The 2024-25 promotion (previous entry) did not replicate; this hypothesis is dead.

Outperformed buy & hold

Test bars

518

Trades taken

109

Strategy return

-10.45%

Buy & hold return

-11.57%

Sharpe (annualized)

-0.46

Max drawdown

-15.97%

Win rate 58.7%Negative risk-adjusted return, consistent with the win rate above. Confirms no exploitable edge for this pair/horizon after realistic spread costs.
GBPUSD — Held out test set, never trained on (1h, horizon 4, submitted 2026-07-10) -- v2 CORRECTED backtest (supersedes entry 3's overlap-inflated Sharpe 0.29). 2024-25 test window. Sharpe -0.71, -3.1% vs buy-hold +5.7%. Was ambiguous before (positive Sharpe but underperformed buy-hold); now unambiguously no edge.

Underperformed buy & hold

Test bars

12,450

Trades taken

982

Strategy return

-3.12%

Buy & hold return

5.65%

Sharpe (annualized)

-0.71

Max drawdown

-4.52%

Win rate 53.6%Negative risk-adjusted return, consistent with the win rate above. Confirms no exploitable edge for this pair/horizon after realistic spread costs.
AUDUSD — Held out test set, never trained on (1h, horizon 4, submitted 2026-07-10) -- v2 CORRECTED backtest (supersedes entry 2's overlap-inflated Sharpe -0.30). 2024-25 test window. Sharpe -1.43, -13.8% vs buy-hold -2.0%. No edge -- confirms the original cheap-screen-false-positive rejection.

Underperformed buy & hold

Test bars

12,453

Trades taken

1,704

Strategy return

-13.84%

Buy & hold return

-2.04%

Sharpe (annualized)

-1.43

Max drawdown

-17.11%

Win rate 53.3%Negative risk-adjusted return, consistent with the win rate above. Confirms no exploitable edge for this pair/horizon after realistic spread costs.
EURUSD — Held out test set, never trained on (1h, horizon 4, submitted 2026-07-10) -- v2 CORRECTED backtest (supersedes entry 1's overlap-inflated Sharpe -0.86). 2024-25 test window. Sharpe -1.80, -14.1% vs buy-hold +6.3%. No edge -- confirms and sharpens the original rejection.

Underperformed buy & hold

Test bars

12,453

Trades taken

1,872

Strategy return

-14.11%

Buy & hold return

6.25%

Sharpe (annualized)

-1.80

Max drawdown

-15.80%

Win rate 54.1%Negative risk-adjusted return, consistent with the win rate above. Confirms no exploitable edge for this pair/horizon after realistic spread costs.
EURUSD — Held out test set, never trained on (1h, horizon 12, submitted 2026-07-10) -- v2 CORRECTED backtest (supersedes entry 0's overlap-inflated Sharpe -0.19). 2024-25 test window. Sharpe -2.52, -19.7% vs buy-hold +6.1%. No edge -- confirms and sharpens the original rejection.

Underperformed buy & hold

Test bars

12,445

Trades taken

1,691

Strategy return

-19.73%

Buy & hold return

6.14%

Sharpe (annualized)

-2.52

Max drawdown

-21.68%

Win rate 50.1%Negative risk-adjusted return, consistent with the win rate above. Confirms no exploitable edge for this pair/horizon after realistic spread costs.