Crash Prediction

ML-powered crash risk monitoring with historical validation and stress testing

This is the early warning system. Check the signal banner below — if it turns orange or red, review the suggested actions before trading. New here? Start with the Getting Started Guide.

Page Guide

Risk Management (dropdown) — Composite exogenous risk score with Tier 1-4 breakdown (Crisis, Sentiment, Narrative, Forward-Looking), individual indicator values, and geopolitical pattern-matching gauge with category scores
Current Signal — Live crash risk level (LOW/MEDIUM/HIGH/SEVERE) with market regime, plain-English interpretation, 7-day signal trend, decision thresholds with suggested actions, potential loss estimates across time horizons, time-to-drawdown estimates, historical crash detection, and 90-day signal history chart
How Accurate Is This? — Signal vs actual outcome table (drawdowns and crash rates by signal level), loss estimate calibration vs targets, and how the model performed during known market crashes
Historical Track Record — 14-year validation across 7 rolling test windows: crash rates by signal, loss estimate calibration, timing accuracy (predicted vs actual days to crash), and stress event detection rates
Stress Test Results — Robustness score (0-100), loss estimate confidence intervals, signal accuracy distributions per level, timing accuracy with percentile bands, stress period analysis, and regime-stratified performance

Loading crash prediction summary...

Risk Management Loading...

Composite Risk Monitor

Here's what's happening across all four risk tiers. A composite score above 50 means external risks are elevated enough to affect the portfolio — check which tier is driving the number.

Composite

Tier 1 (40%)

Crisis Events

Tier 2 (30%)

Sentiment

Tier 3 (20%)

Narrative

Tier 4 (10%)

Forward-Looking

Show individual indicators

Geopolitical Pattern Matching

Why this matters: this model compares current geopolitical conditions to 17 historical crises. A P(event) score above 0.55 means today's pattern resembles past episodes that moved markets.

CLEAR

P(event within 30 days)

Category Breakdown

Alert Thresholds

< 0.55 CLEAR - Normal conditions

0.55 - 0.65 SOFT - Elevated, monitor

> 0.65 HARD - Alert triggered

Model: L2 Classification, Accuracy 0.76
Trained on 17 severity 3+ events

Summary

Signal Trend (Last 7 Days)

How to read this: look at the trend column. A string of worsening arrows means risk is building, even if the signal level hasn't changed yet.

Decision Thresholds

The bottom line: find the row marked "YOU ARE HERE" and follow its suggested action. Each level maps to a specific crash probability and a concrete step to take.

Signal Level

Action

Historical Outcome

What This Means For You

LOW (37% of days)

Normal positioning

5.5% chance of 3% crash in 7d

Market conditions are typical. Stay invested as planned.

MEDIUM (35% of days)

Moderate risk signal

15.3% chance of 3% crash in 7d

Risk is above normal. Historically, tighter stop-losses have helped at this level.

HIGH (21% of days)

Elevated risk signal

29.2% chance of 3% crash in 7d

Nearly 1-in-3 chance of a significant drop. Historically, reducing exposure has been beneficial at this level.

SEVERE (8% of days)

Maximum defense

41.1% chance of 3% crash in 7d

Historically rare and dangerous. In past episodes, defensive positioning has preserved capital.

How Much Could the Market Drop?

Estimated worst-case S&P 500 loss over different time periods. "Likely Worst" means 95% of outcomes are better than this number.

Horizon	Extreme (1%)	Likely Worst (5%)	Magnitude
Loading...

How Soon Could a Drop Happen?

Estimated trading days until a drop of each size. Lower numbers = more urgent warning.

Drop Size	Fastest 1%	Fastest 5%	Fastest 10%
Loading...

Historical Crashes

How the model performed during known market stress events.

90-Day Signal History

What's the point? Seeing three months of signal changes alongside SPY price helps spot whether risk spikes predicted actual drops — or resolved harmlessly.

LOW MEDIUM HIGH SEVERE SPY Price 7d VaR

Model: Machine learning crash prediction model with rolling test validation 2011-2024. VaR calibration: 4.5% exceedance vs 5% target. COVID detected 77%, 2022 bear 61%.