Market Risk Classifier
Your strategy doesn't change.
Your sizing does.
Your alpha, strategy, and infrastructure stay exactly as they are.
Backtest window: 1990-2025 walk-forward. Numbers below are historical, not forecasts.
A daily market-regime classifier with an exogenous input.
The Market Risk Classifier (MRC) classifies each U.S. trading day into a risk shape, delivered before the open. The differentiating input is exogenous environmental data. These signals are not in anyone else's stack, and they are orthogonal to the regime models in the rest of your stack.
The classifier also incorporates SPX and VIX as inputs alongside the environmental signals. The system was walk-forward validated across 1990-2025 (3,270 trading days). Every figure on this page comes from that window unless explicitly labeled otherwise. Full methodology and per-state confusion matrix at mindforge.tech/validation-and-methods.
MRC reports one of these each morning. The rules below reference them by name.
Default state. Markets quiet, VIX median 14. Hedges decay, vol carry earns.
Regime inflection. 92.8% precision (64 of 69), scored per episode on the public record.
Selective heads-up. 76% precision (37 of 49). A confirmation tier, pair with other evidence before sizing.
Rapid vol dislocation. 85% precision (57 of 67). VIX already elevated at firing.
Crisis classifier. 92% precision (12 of 13), 1990-2025. Includes COVID, tariff crisis 2025.
Stop de-risking into recoveries.
Every VIX spike that fades is a false scare. If your hedge overlay scales up on every spike, you pay for protection that does nothing and then sell into the recovery. MRC classifies STABLE on 87% of trading days, and 95% of those STABLE calls were correct (41 of 44 since 2012). When MRC says STABLE, the scare was a false alarm. You hold your position instead of selling into the bounce.
- ›STABLE covers 87% of trading days. When MRC classifies STABLE, the system is saying the current environment does not warrant elevated risk posture. 41 of the last 44 STABLE classifications following a scare were correct: the scare faded and the market recovered.
- ›The value is in what you do not do. Every time a VIX spike triggers a protective response and then fades, the portfolio pays twice: once for the hedge premium and once for selling into the recovery. MRC's STABLE classification tells you which scares are false so you can hold through them.
- ›When STABLE ends, MRC flags the regime change as SHIFTING (93% precision, 2-day median lead before stress shows up in price). That is the trigger to restore full hedge size.
- ›The remaining risk: MRC missing a STABLE-to-crisis transition. In the canonical backtest MRC called 12 of 13 CRISIS events (92% precision, 1990-2025). The single false positive was August 2019 (yuan devaluation), where CRISIS fired but VIX never reached crisis levels.
Cut peak drawdown in half.
Trend strategies make money when regimes hold and lose money when regimes break. The pain is concentrated in a small number of regime-change days. MRC flags those days as SHIFTING, ELEVATED, SHOCK, or CRISIS, naming which severity tier is arriving. You hold full size during STABLE — when trends actually work — and scale down or go flat the moment MRC flags a regime change. Same trend signal. Same execution. Smaller worst-case path.
- ›Trend P&L is generated overwhelmingly during STABLE. Sustained trends require sustained regimes. The conditional policy holds full size during STABLE (87% of days), so the overlay does not give up the bulk of the strategy's P&L.
- ›Annualized return over 1990-2025 walk-forward: 8.82% static vs 6.17% with the regime overlay. Sharpe ratio essentially unchanged (0.63 → 0.62). The reduction in volatility roughly matched the reduction in return.
- ›Episode-by-episode drawdowns: COVID -33.92% → -16.33%. Q4 2018 -10.01% → -7.09%. Aug 2024 -8.49% → -6.87%. Aug 2015 -11.24% → -7.57%. The overlay shrunk every major drawdown in the window.
- ›The overlay is a scaling factor on the position size you were already going to take. Your vol-targeting, multi-asset risk balance, and signal selection are all preserved exactly as they are today.
Close before the regime breaks.
Selling vol earns carry when markets are quiet and breaks expensively when the regime shifts. MRC's SHIFTING state puts a name on the regime shift the morning it is classified, with a scored record behind it. The integration: desks have used the morning SHIFTING classification to lighten short-vol exposure, restoring it once STABLE is classified again. The historical re-entry signal is consistent — 83% of crisis episodes ended with a SHIFTING classification within 10 trading days, which is when carry conditions historically returned.
- ›STABLE-day VIX distribution: median 14.32, with most days between 12.8 and 16.7. Forward 5-day VIX change is essentially zero in median. This is the carry environment short-vol books are designed for.
- ›On SHIFTING days, the median forward 5-day VIX move is -3.3%, but the tails widen sharply (10th percentile -16.4%, 90th percentile +19.8%) compared to STABLE tails. Closing the short-vol book on SHIFTING compresses your exposure to the painful tail.
- ›After a crisis-class state ends, MRC flags SHIFTING within 10 trading days in 24 of 29 historical cases (83%). That is the re-entry signal: the regime has stabilized enough that carry returns.
- ›Late short-vol entry on SHOCK has historically traded into mean reversion. SHOCK classifications happen at already-elevated VIX (median 27.6) and the median forward 5-day VIX move is -19.6%. Wait for the STABLE or SHIFTING that follows.
Reallocate the premium budget.
Long-volatility books bleed premium during STABLE and earn convexity during crises. The same annual budget gets a much better payoff if you spend less during STABLE and more during MRC-flagged crisis states. MRC's CRISIS flagged 12 of 13 events in the canonical backtest (92% precision, 1990-2025), so the scaled-up notional has historically landed where the convexity actually lives. Same dollar budget. Deeper protection where it matters.
- ›The premium is reallocated, not added. Average notional held in the 1990-2025 daily stream was 0.49 versus a constant 1.0 under the static policy. Carry drag (stylized proxy) was cut roughly in half.
- ›Convexity scales up on SHOCK (1.5x) and CRISIS (2.0x). MRC's CRISIS classifications are 12 of 13 (92% precision); SHOCK is 57 of 67 (85%). Each canonical episode is dated and scored in public.
- ›Median VIX at the moment of CRISIS classification was 24.59 across the 9 episodes. SSS classifications historically arrived with VIX still well below crisis-peak levels — meaning the scaled-up notional was purchased before the convexity got expensive.
- ›COVID dominates the expected payoff. 1 of the 9 CRISIS episodes produced a -30.64% SPX drawdown to trough. The other 8 ranged from -0.15% to -6.23%. The modal SSS event is much smaller than COVID. Size accordingly.
- ›The long-convexity payoff numbers above are a directional proxy, not options-pricing. Your actual options book (specific strikes, tenors, vol surface) determines the magnitude on your side.
Tilt factor weights by regime.
Different factors win in different market regimes. Momentum crashes hardest during regime breaks; quality and low-vol earn during stress; value rotates with macro. MRC gives you an exogenous regime input to drive factor tilts — instead of inferring regime from the same prices the factors already use (which is circular). The framework here is a methodology spec; the historical backtest is reproducible on your firm's own factor return series.
- ›Factor returns are regime-dependent. The standard quant portfolio infers regime from the same price data the factors themselves use, which is a circular dependency. MRC provides a regime signal that does not come from the same price stream as your factors.
- ›The momentum-crash episodes in 1990-2025 (Aug 2015, Q4 2018, COVID) all coincided with MRC classifying non-STABLE. The underweight-momentum tilt would have been active across each of those windows.
- ›No published quant backtest in this document because factor-return series are firm-specific. The framework above is documented for reader-side reproduction on Kenneth French Data Library factors or your proprietary series.
- ›The 60-day evaluation provides daily MRC classifications going back to 2012 so your team can merge them with your internal factor backtests directly.
What does the signal cost on its bad days?
Every classifier fires sometimes when nothing follows. If you halve your trend exposure on a SHIFTING that doesn't lead anywhere, you gave up half a position for no reason. We don't bury this. Here's what that looks like.
| State | Activations | False Positives | Precision |
|---|---|---|---|
| SHIFTING | 69 | 5 | 92.75% |
| ELEVATED | 49 | 12 | 75.51% |
| SHOCK | 67 | 10 | 85.07% |
| CRISIS | 13 | 1 | 92.31% |
ELEVATED is selective, not action-grade. 37 of its 49 activations confirmed across 1990-2025 (76% precision); 12 did not. That's why every archetype above treats ELEVATED as a confirmation tier: pair it with other evidence before sizing, not the primary trigger.
The aggregate cost of being wrong is already in the headline numbers. The trend-following backtest gives up ~265 bps annualized (8.82% → 6.17% CAGR) under the regime overlay. That give-up is the cost of all the conditional sizing across 35 years, including every false-positive SHIFTING. The trade is documented in the Section 4 result, not buried.
For day-by-day episode P&L: the validation packet at mindforge.tech/validation-and-methods documents every activation date with its outcome. Reproducible from the daily state series provided in the 60-day evaluation.
Three reasons your team can ship this fast.
The regime signal scales the trades you already make. Your vol-targeting, signal selection, instrument choice, and risk framework all stay exactly as they are today.
Delivered before the U.S. open via API or webhook. No intraday updates. Your risk system reads it once, looks up the sizing rule, applies it before you trade.
The 60-day evaluation delivers daily classifications back to 2012. Your team merges them with your internal backtest infrastructure and reproduces every figure on this page on your own book.
What the pilot actually involves.
The literal: who does what, when, with what data.
- A signed pilot agreement covering confidentiality and use restrictions.
- One point of contact for the weekly briefing.
- Daily MRC classifications by 07:30 AM ET (API or email).
- Full 1990-2025 daily classification series for backfill.
- Validation scripts to reproduce every published statistic on your own infrastructure.
- Weekly research briefs on any episodes during the window.
- Q&A access for integration support.
MRC outputs a daily regime classification, not a buy/sell direction. The classification is a sizing input. All directional and execution decisions remain with the reader's portfolio management and risk governance.