Physics-Derived Market Intelligence
Two systems. Both built from physics. Both predict volatility.
James Lawrence • CTO @ Strykr AI • Forward Deployed Engineer
System 1 classifies volatility regimes on daily data — walk-forward validated across 20 years and 53 assets. System 2 measures structural fragility in real-time order books. Together they form a complete macro risk framework.
72.1%
Regime Accuracy (Daily)
1.93x
Fragility Signal (L2)
Two Systems, One Framework
Strategic bias from regime classification. Tactical timing from structural analysis.
Both systems are built from physical science — one operates on daily price data (macro), the other on real-time order book depth (micro). Neither uses standard financial indicators. Together they answer: "Is the market fragile?" and "Is it about to break?"
System 1 — Regime Classifier (Daily)
13 proprietary physics-derived features. Classifies any asset into one of 6 empirically discovered volatility regimes. Walk-forward validated: 72.1% accuracy, 38.6pp spread, 29,912 out-of-sample signals across 20 years.
System 2 — Structural Fragility (Real-Time)
Computes a proprietary fragility metric from real-time order book depth. Measures how resilient the market's liquidity structure is to perturbations. When it drops, the market is structurally fragile. Validated: 1.93x volatility prediction on institutional-grade L2 data.
Price Data
Daily bars, 53 assets
→
Physics Engine
13 proprietary features
→
Regime Classifier
6 empirical regimes
→
Strategic Bias
Long vol / Short vol
Order Book L2
Real-time depth
→
Proprietary Simulation
Structural analysis
→
Fragility Index
Resilience metric
→
Tactical Trigger
Execute / Hold / Eject
The #1 feature: Primary Structural Signal (59.8% of classifier power)
A proprietary measurement adapted from physics that studies how systems behave near critical thresholds. It detects structural instability before it becomes visible in price or implied volatility. Specification available under NDA.
The Proof
Global Stress Index vs actual VIX (daily, out-of-sample 2020–2026)
Composite Stress Index (red) vs actual VIX (blue dotted). This chart uses the 2020–2026 fold from the daily walk-forward validation. The model was trained only on pre-2020 data.
Global Regime Distribution
What percentage of all 53 assets are in each regime over time
Cyan = assets in Safe Zone (sell vol). Red = assets in Stress Zone (buy vol). Gray = Neutral. When red dominates, systemic vol expansion is likely.
Signal Accuracy
Every SELL VOL signal vs what VIX actually did 90 days later
Green dot = system said "sell vol" and VIX fell (correct). Red X = system said "sell vol" and VIX rose (wrong). Green-to-red ratio = accuracy.
Asset Deep Dives
Per-asset regime history with price overlay
Background shading shows the regime the asset was classified in at each point. Cyan = Safe Zone (sell vol). Red = Stress Zone (buy vol). Gray = Neutral. Green triangles mark SELL VOL signals.
Resolution Analysis
Daily is the proven resolution. Hourly is a research lead.
The daily signal is walk-forward validated across 20 years. The hourly signal showed promise in preliminary (non-walk-forward) testing but does not survive rigorous walk-forward due to limited data history (~2 years via yfinance).
| Resolution | Horizon | Accuracy | Spread | Signals | Status |
|---|
| Daily | 90 days | 72.1% | 38.6pp | 29,912 | Walk-forward validated (7 folds, 20yr) |
| Hourly | 90 days | 69.0%* | 27.6pp | 1,459 | Preliminary (3 folds, 2yr, BUY VOL only) |
Daily short-vol is the core signal
The 7-fold walk-forward on daily data is the primary validated result. Hourly data (limited to ~730 days via public sources) showed a 69% BUY VOL signal in one favorable fold but didn't generalize. With longer hourly history (e.g., from a data vendor), the complementary long-vol edge could be validated — but we won't claim it until it survives walk-forward.
Walk-Forward Validation — Daily Resolution
7 folds, 20 years, retrained from scratch each fold
The model is retrained every 3 years on daily data using only the past. No future information leaks into any fold. This is the gold standard for strategy validation.
| Period | SELL VOL | BUY VOL | Combined | Spread | Signals |
|---|
| 2005–2008 | 56.7% | 83.8% | 71.8% | 43.4pp | 2,903 |
| 2008–2011 | 78.1% | 39.5% | 73.6% | 25.9pp | 5,263 |
| 2011–2014 | 66.8% | 66.7% | 66.8% | 34.9pp | 3,671 |
| 2014–2017 | 57.1% | 55.5% | 56.5% | 37.4pp | 2,774 |
| 2017–2020 | 66.6% | 61.8% | 64.6% | 38.8pp | 4,426 |
| 2020–2023 | 85.3% | 54.0% | 73.0% | 45.8pp | 6,216 |
| 2023–2026 | 71.2% | 53.5% | 62.5% | 44.2pp | 4,659 |
| Aggregate | 72.1% | 60.1% | 67.8% | 38.6pp | 29,912 |
Crisis alpha: the system gets better when it matters
The two highest-accuracy folds are crisis periods — 83.8% BUY VOL before the GFC, 85.3% SELL VOL after COVID. Most quant strategies degrade during crises. This one improves. The physics features detect structural instability that is invisible to price-based models.
System 2 — Structural Fragility
A proprietary fragility metric predicts volatility from order book depth
System 2 computes a real-time "resilience score" from L2 order book data. When the score drops, future volatility increases. The methodology is proprietary. Results below are on institutional-grade Nasdaq L2 data.
| Asset | Correlation | Fragility Ratio | Result |
|---|
| SPY (S&P 500) | -0.42 | 1.93x | Validated — fragility predicts volatility |
| AAPL | +0.17 | 0.98x | No signal — insufficient depth |
| MSFT | +0.01 | 1.06x | No signal — insufficient depth |
| AMZN | -0.16 | 1.05x | No signal — insufficient depth |
| SPY (Placebo) | +0.17 | 0.83x | Signal vanished — not an artifact of volume |
Works on deep markets. Fails on thin ones. Validated by placebo.
The fragility metric produces a 1.93x volatility signal on SPY (the deepest order book in equities). It fails on single stocks where depth is insufficient. A placebo test (randomized book structure) destroyed the signal, confirming it captures genuine structural properties — not just aggregate volume. Applicable to any deep liquid market: indices, FX, crypto majors. Methodology available under NDA.
What Drives the Signal
Feature importance — what drives the classifier
13 proprietary physics-derived features. The top 3 account for 78.0% of total predictive power (averaged across 7 walk-forward folds). Feature names are obfuscated; specifications available under NDA.
Short-Vol Accuracy by Asset Class
Regime Transitions
Last 60 days — which assets are transitioning
Each cell = one asset on one day. Color = regime. When rows shift color simultaneously, a systemic regime transition is in progress.
Validation
Head-to-head against the industry standard
Same 53 assets, daily resolution. Same binary question: "Will realized vol be higher or lower 90 days from now?" Our system uses 7-fold walk-forward; GARCH uses expanding window.
| Metric | GARCH(1,1) | This System (v2) |
|---|
| SELL VOL accuracy | 58.4% | 72.1% |
| BUY VOL accuracy | — | 60.1% |
| Best-worst regime spread | — | 38.6pp |
| Out-of-sample signals | — | 29,912 |
| Test methodology | Single split | 7-fold walk-forward |
| Test period | 2020–2026 | 2005–2026 (20 years) |
| Assets / classes | 53 / 15 | 53 / 15 |
Known limitations
72% directional accuracy does not guarantee P&L — short-vol strategies carry asymmetric tail risk. All results are historical simulations, not live trading. BUY VOL accuracy (60.1%) is weaker than SELL VOL. 90-day horizon is medium-term — not suitable for intraday or HFT. Execution and position sizing remain open problems.
Technical
Built to run, not just to publish
| SYSTEM 1 — REGIME CLASSIFIER |
| Proprietary features | 13 (physics-derived, proprietary) |
| Discovery method | Gaussian Mixture Model on 381,264 bars |
| Classifier | Gradient-boosted ensemble + isotonic calibration |
| Regimes | 6 (discovered empirically, not predefined) |
| Validation | 7-fold walk-forward on daily bars (2005–2026) |
| Out-of-sample signals | 29,912 across 53 assets and 15 classes |
| Anti-overfitting | Retrained 7x; signal holds in every fold (56–85%) |
| SYSTEM 2 — STRUCTURAL FRAGILITY |
| Method | Proprietary simulation on L2 order book depth |
| Metric | Proprietary fragility index (resilience score) |
| Signal | -0.42 correlation with future volatility (SPY) |
| Fragility ratio | 1.93x (fragile vs resilient regimes) |
| Placebo validated | Signal vanishes on randomized book data |
| Applicability | Deep liquid markets (indices, FX, crypto majors) |
| SHARED |
| Stack | Python, NumPy, pandas, SciPy, scikit-learn, gradient boosting |
FAQ
Is this just overfitting? +
No. The system is walk-forward validated across 7 separate 3-year periods from 2005 to 2026, retrained from scratch each fold. The combined accuracy (67.8%) never drops below 56.5% in any fold. The 38.6pp regime spread held stable across 20 years of different market conditions — bull markets, crises, low-vol eras, rate cycles.
Why not just use VIX? +
VIX is a reaction measure — implied vol as priced by the options market right now. This system is predictive — it identifies structural fragility that precedes the VIX spike, not the spike itself. In testing, the Stress Index led VIX moves by days to weeks.
What about transaction costs? +
The backtest includes 5bps transaction costs per trade. The strategy is medium-term (90-day horizon) with low turnover (18% time in market), so costs are not the primary driver of P&L. The main risk is the asymmetric payoff of short volatility, not slippage.
Is this a black box? +
No. Every feature has a specific physical interpretation and measures a distinct structural property of market dynamics. We know exactly why the model classifies each regime, unlike deep learning approaches. Full feature specifications available under NDA.
Can a fund replicate this from the website? +
No. The website shows results and obfuscated feature names — not formulas, thresholds, window sizes, or classifier hyperparameters. The 13 feature definitions, training methodology, and model configuration are proprietary.
Why physics instead of standard quant factors? +
Standard factors (momentum, mean-reversion, carry) describe what happened. Physics features describe the structural state of market dynamics — properties orthogonal to traditional signals that price-based indicators miss entirely.
Which asset classes work best? +
Bonds (94.8%), emerging markets (90.8%), and European equities (85.5%) show the strongest classification accuracy. Commodities and FX are in the 65–72% range. Credit and single-name tech equities are weakest (~49%). The system works best on assets with well-defined volatility cycles and sufficient history.
How would a fund deploy this? +
Three options: (1) Overlay — run the classifier daily on a vol-selling book and skip trades when the system flags Stress Zone, reducing tail risk. (2) Signal generation — feed regime labels into an existing risk model as an additional factor. (3) Standalone — sell vol when Safe Zone, buy vol when Stress Zone, flat otherwise. The 90-day horizon suits systematic macro and vol-arb desks.