AlgoSignal Lab

24 Validated Signals + research100 Sweep Findings — Updated 2026-04-28

• ✓ 204 Gate C candidates (H1 QC-mode sweep) across 31 symbols — Sharpe 0.5–3.0, QC-mode, ≥3 symbols
• ✓ 24 QC-validated signals across 17 instruments — 7 Tier A, 12 Tier B, 5 Tier C
• ✓ USO-233 research-backed batch: 5/5 PASS — vol_adjusted labeling delivers +0.6–1.5 Sharpe improvement over fixed-threshold on carry-aligned pairs
• ✓ New signals: EURNZD +0.562 (USO-230) · XAUUSD Ensemble +0.538 (USO-222) · GBPNZD Ensemble +0.684 (USO-224)
• ✓ research100 sweep (USO-254): 86 features · 5,040 experiments · 35 instruments · 144 configs — 3,457 extended-gate PASS (68.6%) · mean R/DD 50×, median 31× — highest-performing sweep to date
• ↗ Top uncapped candidate: CADCHF r053 LightGBM simple — R/DD 832.8 · Return 82.2% · MaxDD 0.10% · 821 trades · WR 85.1%
• ✗ P1–P5 retired — all QC FAIL, superseded by enhanced 65-feature model set
• paper_trader_v2.py live on Powerserver · models at /workspace/fx_ml/production_models/manifest.json (USO-204)

24 Validated Signals — Full Portfolio (USO-234 + USO-230 + USO-233)

• Is the test data out-of-sample? Yes — 8-fold walk-forward, 2021–2024 OOS windows. Strict chronological split enforced in code. No look-ahead bias (USO-109 audit passed).
• How many strategies are active? 24 signals total: 7 Tier A (high confidence), 12 Tier B (active — incl. 5 research-backed), 5 Tier C (caution). See Portfolio tab.
• Best production candidates? CADCHF LightGBM (+39.40, 8/8 folds), AUDCHF XGBoost (+34.48, 8/8 folds), USDCHF XGBoost (+28.12, 8/8 folds). CHF pairs dominate Tier A.
• Research-backed additions (USO-233)? 5 vol_adjusted signals (Barroso & Santa-Clara 2015): USDJPY CatBoost +2.716, USDCAD Logistic +2.052/+1.995, USDJPY LightGBM +1.771, NZDUSD RF +1.640. Credible Sharpes — no artifact inflation.
• New signals from USO-230/222/224? EURNZD XGBoost +0.562 (4/7 folds), XAUUSD Ensemble +0.538 (8/8 folds — genuine Gold signal), GBPNZD Ensemble +0.684 (6/8 folds).
• What happened to P1–P5? All retired — QC FAIL. Replaced by enhanced 65-feature P6–P13 model set (USO-202).
• CAUTION signals? P10 CADJPY, P12 EURGBP (QC Sharpe pending), P15 XAUUSD CatBoost (1/8 folds — artifact). P23-P25 new Tier C — reduce position sizing to 50%.

Why Tree Models Beat LSTMs on Forex 1-Min Data

High noise-to-signal ratio on 1-min bars makes sequential models prone to overfitting spurious patterns. GBDTs (LightGBM, XGBoost) with tabular feature engineering match or exceed LSTM performance with 3–10× faster training and interpretable SHAP attributions.

• LightGBM is the default choice: leaf-wise growth + native NaN handling + fast histogram splits
• XGBoost preferred for small datasets or when regularization needs fine-tuning
• 1-min data requires at least 2 years of history for reliable walk-forward splits
• Confidence thresholding at 0.5 filters low-conviction signals and improves net Sharpe

63-Feature Engineering Pipeline for EURUSD H1

Built from OHLC-only MT5 data (no volume, no order book). Features grouped into 7 categories covering price microstructure, trend, momentum, volatility, volume proxy, and session timing.

• Session dummies (London open, NY open, overlap) capture intraday regime shifts
• Multi-timeframe returns (ret_1, ret_5, ret_15, ret_60) encode mean-reversion vs momentum
• Bollinger Band position (bb_pos) is a strong confidence filter for entries
• ATR-based sizing (atr_7, atr_14) enables dynamic position sizing in risk layer
• Stochastic K/D adds short-term overbought/oversold context beyond RSI

EURUSD H1 6M LightGBM Results

Walk-forward validation over 6 months of H1 data with 4-fold time-series splits. Best run (87e9fdae) achieved net Sharpe 0.418, 3.24% return, profit factor 1.03 — marginal but consistent positive edge confirmed across 3 of 4 folds.

• Fold 3 is strongest: Sharpe 3.87, +9.4% net return in isolation
• Fold 2 is weakest: Sharpe −3.02 — market regime change visible
• Average 1,453 trades per full run provides statistical significance
• 90.1% bar coverage (signals generated for most bars above threshold)
• Net vs gross gap is small: spread cost is manageable on H1 EURUSD

Trade Frequency Optimization & Position Sizing

Optimal signal delivery cadence is H1 with 3–6 signals per session. Kelly criterion-based position sizing with 25% cap prevents over-leverage. Fixed fractional sizing (1–2% risk per trade) recommended for subscribers.

• H1 provides best signal-to-noise tradeoff vs 1-min (too noisy) or D1 (too infrequent)
• ATR-based stop-loss placement reduces max drawdown by ~40% vs fixed-pip stops
• Confidence threshold of 0.55–0.60 further improves Sharpe at cost of reduced coverage
• Kelly fraction capped at 0.25 to control ruin risk for subscriber accounts

Statistical Validation & Reproducibility Risks

Key risks identified: look-ahead bias in feature calculation, overfitting to specific H1 regime, and local vs QuantConnect environment variance (execution, spread, data source differences).

• All features use strict lag-1 alignment — no future data leakage confirmed
• QuantConnect replication (USO-22/23) required before production signal launch
• Recommended: combinatorial purged cross-validation (CPCV) for final validation
• Monte Carlo simulation of drawdown distribution needed for subscriber risk disclosure

research100: 86-Feature Sweep Across 35 Instruments — Highest-Performing Sweep to Date

Expanded feature set (86 features, up from 65) swept across 35 instruments × 144 configs = 5,040 experiments. Extended gate (no R/DD cap) reveals exceptional candidates: 3,457 pass (68.6%), mean R/DD 50×, median 31×. Standard gate (R/DD 0–5) yields 514 candidates (10.2%).

• Top uncapped candidate: CADCHF r053 LightGBM simple — R/DD 832.8 · Return 82.2% · MaxDD 0.10% · 821 trades · WR 85.1%
• Best return (standard gate): EURJPY CatBoost triple_barrier — R/DD 4.98 · Return 46.8% · 18,074 trades · WR 77.0%
• LightGBM: fewest extended-gate passes but highest mean R/DD (74.6) — quality over quantity
• XGBoost: middle ground (53.6 avg R/DD); CatBoost: most passes, lowest mean R/DD (32.4)
• Top symbols by pass count: GBPNZD, EURJPY, CHFJPY, USDCAD, EURGBP, EURNZD, CADCHF, GBPCHF, AUDNZD, AUDCHF
• Extended pass rate of 68.6% vs 10.2% std gate confirms 86-feature set captures strong directional bias in many instruments — QC validation pipeline queued for top candidates

Start Paper Trading Immediately — Close the Track Record Gap

Every revenue, trust, and distribution initiative is blocked without a live track record. The model already clears the minimum performance bar (Sharpe ≥ 0.35 OOS, DD ≤ 20%) but has zero live weeks. Deploy EURUSD H1 LightGBM to MT5 demo or QC live paper mode this sprint.

• Connect to MyFXBook/FX Blue on day one — verification starts from first trade
• Build Telegram bot in parallel — distribution channel ready when 12-week minimum is reached
• Unblocks: MQL5 listing decision (USO-35), signal subscription launch (USO-11), Collective2 evaluation (USO-121)

Verify the Sharpe +0.42 Result Before External Use

USO-24 explicitly flagged this as a hypothesis, not a verified finding. Look-ahead bias from a scaler fitted on the full dataset is the single most common silent inflator — if real OOS Sharpe is materially lower, building marketing on it is catastrophic when live results diverge.

• Run USO-24 checklist: confirm per-fold scaler scope, label direction, H1 resampling alignment
• Compute DSR (at N=50, a Sharpe of 0.42 is at the expected max by chance)
• Re-run with explicit EURUSD H1 data; report mean ± std Sharpe across ≥5 walk-forward folds

Implement the 3-Strategy Experiment Speedup Stack

Brute-force grid search will take months to reach 2,000+ experiments (USO-90). USO-92 identified a combined 70–85% wall-clock reduction available in ~2 weeks via SHAP pre-screening + Optuna TPE + Ray Tune.

• Week 1 — SHAP pre-screening gate: discard bottom 50–60% of features (626 → ~250), ~30–50% savings/experiment
• Week 1 — Optuna TPE + HyperbandPruner: reaches top-10% configs with 10–20× fewer trials
• Week 2 — Ray Tune distributed: orthogonal throughput multiplier across both servers

Fix the QC Cost Model Across All Backtests

QC OANDA default = zero spread + zero slippage. EURUSD actual = ~1.5–2.3 pip round-trip. This already caused gross Sharpe +2.46 → negative net Sharpe at default threshold (USO-13).

• Add ConstantSlippageModel(0.000015) + SpreadSlippageModel() to every QC algorithm config
• Mandatory spread stress gate: positive Sharpe at 3.0 pip before any result is reported
• Add cost-parameter block to all backtest reporting templates

Build the Multi-Strategy Portfolio

All product revenue currently depends on a single EURUSD H1 LightGBM model. USO-27 identified 13 strategies; 4 Tier 1 strategies each require ~1–2 days to implement. Target: 4 uncorrelated strategies (ρ < 0.3) → portfolio Sharpe 0.7–1.2 vs. 0.42 single-strategy.

• Implement volatility-conditioned sizing, quantile stops, microstructure score, Fourier features (USO-27 Tier 1)
• Each strategy must pass USO-24 checklist individually before portfolio combination
• Directly supports the Pro $79/mo tier and provides a competitive moat

Lock the Go-to-Market Phase Sequence

• No MQL5 listing at MVP — require live track record first
• Collective2 at Phase 3. Telegram bot before MT5 EA

Adopt One Canonical Sharpe Metric + Experiment Registry

Performance is currently referenced inconsistently across documents (gross vs. net, H1 vs. 1-min, single-split vs. WFV mean). This is an internal confusion and external credibility risk.

• Canonical: annualised_sharpe = mean(daily_returns)/std(daily_returns) × √252, net of 2.3-pip round-trip, as walk-forward mean ± std across ≥5 folds
• Version-controlled experiment registry: model_id, run_date, instrument, timeframe, N_folds, mean_sharpe, std_sharpe, N_variants, DSR, spread_assumption
• All external communication uses only canonical walk-forward net OOS Sharpe

5 Critical Caveats Before Productising the Shortlist

• No transaction costs in backtest — all Sharpe figures are gross of spread, commission, and slippage. Apply USO-44/USO-49 cost corrections; expect 20–40% Sharpe degradation on FX majors, higher for Energy/Metals.
• Single-period backtest — the sweep covers one fixed historical window. Without walk-forward OOS or CPCV (USO-24), Sharpe values have unknown generalization confidence.
• USDJPY look-ahead audit required — 292 of 564 Sharpe > 5 entries concentrate on USDJPY. Anomalous across all models and horizons; must be audited for feature leakage before any live use.
• Win rate ≠ profitability — several high-Sharpe entries have 35–40% win rates. Verify R:R structure is compatible with USO-13 Kelly 0.25× / ATR exits 2:1 R:R framework.
• LightGBM dominance risk — 5 of 8 shortlisted signals use LightGBM. If it systematically overfits to a backtest regime, multiple signals will fail simultaneously. Retain XGBoost entries (#3 GBPUSD, #8 AUDUSD).