Hypotheses

EURUSD 4h-ahead direction from basic technical indicators

Predict the sign of EURUSD's return 4 hours ahead using SMA trend, momentum, rolling volatility, RSI, and recent high/low range as features, on hourly candles.

AUDUSD 4h-ahead direction (cheap screen)

Tests whether AUDUSD's recent momentum, volatility, RSI, and distance from its 20-period high/low can predict which direction price moves over the next 4 1h bars.

Full validation (corrected v2 backtest): Sharpe -1.43, -13.8% return vs buy-hold's -2.0%, 53.3% episode win rate. The cheap-screen promotion was a false positive -- doesn't hold up under real testing.

NZDUSD 4h-ahead direction (cheap screen)

Tests whether NZDUSD's recent momentum, volatility, RSI, and distance from its 20-period high/low can predict which direction price moves over the next 4 1h bars.

Below the cheap-screen bar -- rejected before spending the expensive full validation budget.

USDCAD 4h-ahead direction (cheap screen)

Tests whether USDCAD's recent momentum, volatility, RSI, and distance from its 20-period high/low can predict which direction price moves over the next 4 1h bars.

Below the cheap-screen bar -- rejected before spending the expensive full validation budget.

GBPUSD 4h-ahead direction (cheap screen)

Tests whether GBPUSD's recent momentum, volatility, RSI, and distance from its 20-period high/low can predict which direction price moves over the next 4 1h bars.

Full validation (corrected v2 backtest): Sharpe -0.71, -3.1% return, buy-and-hold did better (+5.7%). Pre-correction this read as technically positive (Sharpe 0.29) -- the fix clears up the ambiguity into a clean rejection.

USDJPY 4h-ahead direction (cheap screen)

Tests whether USDJPY's recent momentum, volatility, RSI, and distance from its 20-period high/low can predict which direction price moves over the next 4 1h bars.

Below the cheap-screen bar -- rejected before spending the expensive full validation budget.

USDCHF 1D horizon=20, central-bank sentiment (LLM-scored Fed/SNB statements)

LLM-scored hawkish/dovish stance from Fed and SNB policy statements (cbank_sentiment_diff, cbank_sentiment_diff_chg_63d) on the same USDCHF/1D/h=20 combo as the rejected carry features below. Full validation: Sharpe 0.44 on 2024-25 (+6.1% vs buy-hold's -5.3%, 100% bars in-market, 5 trade episodes) but Sharpe -0.66 on the 2021-22 robustness window (-7.8% vs buy-hold's +4.9%) -- failed robustness, same pattern as every 1D/h=20 carry candidate below. The one place an LLM has a real comparative advantage in this pipeline (reading text), tested the same way as everything else -- and still didn't hold up.

USDCHF 1D horizon=20, carry (rate_diff) features

Backtest shows Sharpe 0.32 (+4.5% vs buy-hold -5.3%), but the model took exactly 1 trade episode for the entire 500-bar test window (in-market 100% of the time, one static position held throughout). Not a timing strategy -- a single directional bet that happened to be right. The '100% win rate' is trivially true with n=1 and should never be quoted as a real result.

NZDUSD 1D horizon=20, carry (rate_diff) features

Promising on 2024-25 (Sharpe 0.77, +13.6% vs buy-hold's -8.4%, 60 episodes) but did not replicate on 2021-22: val AUC 0.31 (far worse than random, majority-class baseline 0.65), Sharpe -0.002 (flat), max DD -17.9%. Same lucky-window pattern as AUDUSD (1D, h=12) on 2026-07-10.

GBPUSD 1D horizon=20, carry (rate_diff) features

The strongest cheap-screen hit of the whole carry batch -- rate_diff_chg_63d cleared the IC bar easily (+0.13). Full model didn't replicate it: val AUC 0.513 (barely above chance), backtest Sharpe -0.70 (-8.7% vs buy-hold +4.7%, max DD -14.2%, in-market 92% of bars). A clean example of a cheap-screen promote not being a confirmed edge, even with a mechanistically plausible feature.

USDJPY 1D horizon=20, carry (rate_diff) features

rate_diff_chg_63d was the best feature here too (IC -0.14), but val AUC came back 0.454 -- worse than random. Backtest Sharpe -0.04 (-1.7% vs buy-hold +9.5%, max DD -12.2%, in-market 95% of bars).

EURUSD 1D horizon=20, carry (rate_diff) features

Underperforms buy-hold with weak Sharpe: 0.18 (+1.8% vs buy-hold +6.1%). Doesn't beat a coin flip in any meaningful sense.

USDCAD 1D horizon=20, carry (rate_diff) features

Negative Sharpe -0.19 (-2.1% vs buy-hold +5.1%).

AUDUSD 1D horizon=20, carry (rate_diff) features

Sharpe 0.14 (+1.6% vs buy-hold -3.2%) -- only ahead because buy-hold itself was negative that window, not a real edge.

USDCHF 4h horizon=24 (technical indicators)

Old (v1, overlapping-return) backtest showed Sharpe 1.39 / +118% -- entirely an artifact of the overlap bug (val AUC 0.44, worse than random). Corrected v2 backtest: Sharpe -0.13 (2024-25) and -0.42 (2021-22). A second, independent USDCHF timeframe that did NOT replicate the confirmed 1h/h=4 edge.

AUDUSD 1D horizon=12 (technical indicators)

Promising on 2024-25 (Sharpe 1.09, val AUC 0.63) but did not replicate on 2021-22 (Sharpe -0.46, val AUC 0.46). Likely a lucky window -- the same failure mode later seen in the 1D/h=20 carry batch.

USDCHF 1h h=4: exclude NY session hours (16-21 UTC)

A session breakdown of the confirmed model showed it both trading most (64% in-market vs. its usual ~35%) and losing money during NY hours specifically. Excluding NY hours looked great on 2024-25 (Sharpe 0.83 -> 1.68, +8.2% -> +14.9%) but failed robustness on 2021-22 (Sharpe 1.98 -> 0.54, +22.7% -> +5.0%, max DD got worse -4.0% -> -5.1%). The session pattern was specific to one window, not persistent -- this filter is NOT applied to the live model.

Liquidity sweep-and-reject fade (cross-pair)

"Everyone does X, so X should reverse" is a pattern, not a mechanism -- rejected at the cheap-screen stage before spending full-validation budget. 185 sweep events at 4h (excess return vs. baseline t=0.78, not significant) and 51 events at 1D (t=-1.33, wrong sign). See scripts/scan_sweep_reversal.py.

NZDUSD 4h horizon=12 (technical indicators)

Rejected at the cheap-screen stage -- never reached full validation.