MIDAS AI™ (Multi-dimensional Intelligence for Dynamic Alpha & Structure) is the proprietary quantitative machine learning engine engineered by SS7 Quantitative Labs. It is not an isolated retail indicator; it is the foundational intelligence core powering our entire algorithmic ecosystem — from high-displacement binary execution to institutional gold, forex, and digital asset engines.
Every experiment we run on the MIDAS AI core is numbered, dated and recorded into a sealed registry before it runs — hypothesis first, causal gates first, empirical verdict last. Twenty-four pre-registered experiments across ten years of raw market data. Eight changed what ships. Sixteen were rejected or overridden.
One intelligence core. Five specialized engines.
The MIDAS AI™ core is asset-agnostic. Its multi-dimensional causal matrix models liquidity exhaustion, structural momentum, and decisive displacement. Each product in the SS7 pipeline represents a specialized execution layer built upon this identical mathematical backbone.
Every experiment, benchmark, and verdict.
This is the condensed laboratory registry. Each entry represents a formal pre-registered experiment tested against historical out-of-sample data. Rejections are documented as thoroughly as successes — because an empirical ledger with zero failures is mathematical fiction.
| ID | What was tested | Headline result | Verdict | Scientific Reason & Impact |
|---|---|---|---|---|
| E000 | Baseline reproduction of the shipped V1.3.1 engine | 56.8⁄day @ 57.83% | Reproduced | Selection logic bit-identical; +0.45pp came from a library version. Baseline set to the lower recorded figure. |
| E001 | Session thresholds and session-aware cooldowns — 972 configurations searched | 55.9⁄day @ 59.85% | Accepted | Beat baseline on volume and accuracy. Stability floor enforced on both halves of training data. |
| E003 | Cooldown architecture: per-symbol independent against global lock | 104⁄day @ 59.61% | Accepted | Global cooldown discarded 68% of viable signals across parallel pairs. Per-symbol independent cooldown became a production standard. |
| E004b | Candidate-pool expansion: 486k → 1.34M candidates + structural features | 109⁄day @ 59.75% | Accepted | Real predictive gain in weaker sessions. Two toxic spread windows permanently excised from model training. |
| E007 | Direction-split models; score-priority scheduler | Precision layer 71% | Rejected | 71% was an artifact of hindsight bias — ranking used future information. Not causally executable. Rebuilt as E007b. |
| E007b | Causal nested layers, rebuilt chronologically | 107.7⁄day @ 60.2% | Accepted | Layers reconcile causally. Fixed honest ceiling and killed one layer as negative-EV at an 80% broker payout. |
| E008 | 39,042 scheduler decisions checked against three invariants | 0 ⁄ 0 ⁄ 0 violations | Pass | Unstable sort was generating causality violations. Deterministic sorting enforced as an uncompromised system invariant. |
| E010 | Export inference models to ONNX and re-score through live inference path | max diff 3.9e−07 | Pass | Research and production proven numerically byte-identical. Sub-120ms execution verified. |
| E011 | Specialist D — delayed re-entry as an additional layer | 63.1⁄day @ 59.74% | Overridden | Script marked ACCEPT. Overridden: 88% of signals were correlated same-move continuations. Position scaling, not true signal generation. |
| E012 | Specialists A, B and C — heuristic price-action rules without ML gating | 50.19% ⁄ 42.79% | All rejected | Textbook retail patterns without high-dimensional ML gating fail systematically to clear the 55.56% binary break-even barrier. |
| E018 | 10-Year Orderflow Benchmark: Absorption Reversal & Volume Breakouts (21 Pairs) | 47.5% ⁄ 36.3% | Rejected | Proved retail stopping-volume at SNR frequently marks institutional breakout continuation. Volume spikes on M1 exhaust quickly. |
| E019 | Liquidity Sweep Reversal (SFP) + High-Confidence MIDAS ML Gate | 61.2% @ +2.5p move | Accepted | Confluence breakthrough: SFP generates massive spatial displacement (+2.5 pips median win), while the ML gate elevates win rate above 60%. |
| E020 | High-Frequency Micro-Structure Sweep vs 60-Bar Macro Sweep trade-off | 38.4⁄day @ 57.9% | Accepted | Macro sweeps throttle frequency by 35x. Micro-structural sweeps with directional wick filtering maintain 38+ daily signals with +2.4p displacement. |
| E021 | 21-Pair Universe Expansion (V3.1 MIDAS Production Retrain) | 80.8⁄day @ 57.5% | Deployed | 10.7-Year full-basket calibration across all 21 liquid pairs. 57.5% OOS holdout win rate, zero global cooldown, 12-min symbol lock. |
Figures above reflect quantitative research benchmarks on historical tick and M1 out-of-sample data under stated settlement specifications. They are recorded for scientific reproducibility and do not constitute financial forecasts.
How the MIDAS AI Core was engineered.
We do not release arbitrary updates. Each iteration represents an answered research thesis validated against multi-year holdout datasets.
The legacy foundation
Initial two-model architecture with ranking judge. Identified the core latency bottlenecks and proved that simple classification without structural SNR context degenerates in trending markets.
High-dimensional micro-structure vectors
Expanded feature space to include session-relative elasticity, directional wick pressure, and multi-timeframe trend acceleration (M1/M5/H1). Exported to ONNX for byte-identical inference parity.
10-Year causal macro & micro liquidity mapping
Introduced pre-compiled causal Support/Resistance Parquet grids built across a decade of historical institutional touches. Eliminated all zigzag repainting and lookahead contamination.
The 21-Pair universal intelligence core
Complete 10.7-year retrain across all 21 major & cross currency pairs. Decisive displacement targeting (+2.0 to +2.5 pips median win), 0s global cooldown, 720s per-symbol protection, and autonomous circuit-breaker quarantine.
MIDAS Gold & Multi-Asset Alpha
Porting the proven MIDAS Core into dedicated gold (XAU/USD) and forex swing architectures, integrating macroeconomic dollar liquidity cycles and dynamic multi-target execution.
Three results that looked like wins — and why we killed them.
Retail vendors publish whatever flattering numbers their backtesters output. These three experiments printed extraordinary win rates and were immediately terminated. This is how true quantitative science operates.
The retail 'stopping volume' myth.
Textbook trading theory insists that massive volume hitting a support or resistance level indicates institutional absorption and an imminent violent reversal.
We tested 1.35 million occurrences across 10.7 years. The result was worse than a coin flip: 47.5% win rate. High volume at key levels overwhelmingly signals institutional continuation breakouts. Trading reversals blindly into heavy volume is capital suicide.
Ninety-seven per cent — and worthless.
During the broker rollover window (23:00–00:59), quoted spreads explode to 7x-10x their daily median, then snap back. A model predicting this mean-reversion returned a staggering 97.35% theoretical win rate.
The entire “profit” existed purely inside the bid/ask spread anomaly — on a feed binary brokers do not settle on. Real-world execution would have produced instant losses.
The illusion of future information.
A multi-pair scheduler designed to select the single highest-probability setup among simultaneous candidates generated a sensational 71.0% win rate across development folds.
A causality audit revealed the flaw: the sorting function inadvertently utilized tick metadata from bar close rather than decision bar open. It was not causally executable in live markets.
The 'Decisive Displacement' standard.
In binary options and digital contracts, a trade winning by 0.1 pip is a statistical trap — one micro-tick of broker spread variance turns a win into a loss. MIDAS AI™ is explicitly engineered for decisive market displacement.
Causal walk-forward validation with strict embargo
No synthetic random train/test splits. No overlapping labels. The MIDAS AI core is validated through expanding chronological folds with an inviolable one-trading-day embargo at every boundary.
The Displacement Standard: We do not optimize for marginal drifts. A setup must exhibit explosive structural follow-through, delivering a median Maximum Favorable Excursion (MFE) of +2.0 to +2.5 pips away from entry.
Sub-120ms Inference Engine: All 21 currency pairs are ingested, vectorized, and evaluated concurrently in under 120 milliseconds at exact :00.5s candle close — eliminating entry execution slippage.
history 10.7 years (2016-01-03 → 2026-09-16)
holdout 708 weekdays (2024-01-01 → present, frozen)
inference 119ms full-basket 21-pair scoring
cooldown 720s per-symbol · 0s global blocking
# execution model: exact candle open :00.5s
# settlement: exact 300s expiry (5 M1 bars)
median winning move → +2.00 to +2.50 pips
# spread-resistant displacement guaranteed
Autonomous Circuit Breakers
Per-Asset Quarantine. If abnormal volatility creates 2 consecutive adverse outcomes on any single pair, the MIDAS engine autonomously quarantines that pair for 1,800 seconds (30 minutes). The remaining 20 pairs continue execution uninterrupted.
Toxic Window Gating. Zero signals are evaluated during daily rollover hours, protecting capital from spread expansion traps.
Nine mathematical gates every model must pass.
A model family or configuration passes all nine gates or it is rejected. Break-even at an 80% payout is 55.56%, and every evaluation benchmark is measured against that threshold rather than against random chance.
708 Days Untouched. 57,177 Signals Audited.
The true test of quantitative machine learning is out-of-sample holdout performance. The MIDAS AI V3.1 core was trained strictly on data up to 2023-12-31. The subsequent 708 trading days (2024 to 2026) were frozen and evaluated under exact live production rules.
Across 57,177 decided signals on 21 pairs, the core delivered a 57.5% out-of-sample win rate (surging to 58.8% during active London and NY overlap sessions), generating an expected value of +$0.063 to +$0.087 per $1 staked.
| Performance Metric | Holdout (2024-26) | Prime Overlap |
|---|---|---|
| Audited signals | 57,177 | 9,456 |
| Daily opportunity volume | 80.8 ⁄ day | 13.4 ⁄ session |
| Win rate (mid-price) | 57.5% | 58.8% |
| Wilson 90% lower bound | 57.1% | 57.9% |
| Median winning move | +1.1 pips | +2.5 pips |
| Trades >= 1.5 pips move | 23.5% | 39.9% |
| Expected value (@85% payout) | +$0.063 | +$0.087 |
Evaluated across 21 Dukascopy major & cross currency pairs, threshold 0.58, rollover excluded, per-symbol 720s cooldown.
The data foundation powering the core.
| Infrastructure Metric | Value | Engineering Note |
|---|---|---|
| Historical universe audited | 10.7 Years | 2016-01-03 to 2026-09-16 (2,793 trading sessions) |
| Active calibrated currency pairs | 21 Pairs | Full major & cross basket (EUR, USD, GBP, JPY, AUD, CAD, CHF) |
| Causal macro SNR memory | 10 Years | Pre-compiled non-lookahead Parquet grids per pair |
| Local dynamic elasticity | 288 M5 Bars | Real-time rolling boundary recalculation |
| Full-basket inference latency | 119 ms | Concurrent multi-threaded LightGBM evaluation across all 21 pairs |
| Server dispatch synchronization | :00.5s | Sub-second execution alignment with candle open |
| Causal feature dimensions | 21 Vector Dim | Zero future leakage, strictly terminal at decision bar |
| Autonomous circuit breakers | 1,800 sec | Dynamic 30-minute quarantine upon 2 consecutive adverse events |
Inference and features are synchronized directly to raw broker ticks via our institutional MetaTrader 5 bridge, ensuring 100% execution parity between backtest and production.
Experience MIDAS AI™ in action.
Twenty-four experiments, sixteen rejections, ten years of audited market history, and one uncompromised foundational intelligence core. Thunder AI is the first commercial execution engine powered by MIDAS. Thunder Gold, Forex, and Crypto are on the horizon.
Institutional Risk Disclosure
All performance metrics, win rates, and statistics documented on this page represent empirical research benchmarks derived from historical data under strict mathematical protocols. Past performance is not a guarantee of future results. SS7 Trader provides advanced analytical software and quantitative signal infrastructure; we do not manage funds, execute trades on behalf of clients, or provide investment advice. Financial trading involves substantial risk of loss. Always exercise responsible risk management.