Surviving the March 2020 Regime Break
How an HMM-based regime detection system automatically adapted to the COVID volatility event, limiting drawdown to 8.2% while static-parameter competitors suffered 34%.
The Problem: Static Parameters in Non-Stationary Markets
The trading system was trained on 2018–2019 data — a predominantly trending regime with low VIX. Parameters were optimised for that environment: wide stop-losses, aggressive position sizing, momentum-following entries.
When the COVID-19 sell-off began on February 20, 2020, the VIX spiked from 14 to 82 in three weeks. The system's static parameters — calibrated for a regime that no longer existed — produced a 34% drawdown on backtests run without regime adaptation.
Drawdown Comparison: Adaptive vs Static
Figure 1: Maximum drawdown trajectory, Feb–April 2020.
Detection: 3-State Hidden Markov Model
The system uses a Hidden Markov Model with Gaussian emission distributions, trained on rolling 252-day windows of realised volatility, return autocorrelation, and VIX term structure slope.
State 1: Trending
σ < 15% · ACF(1) > 0.1
Momentum signals are valid. Full position sizing. Standard stop-loss widths.
State 2: Mean-Reverting
σ 15–25% · ACF(1) < -0.05
Reduce position size by 40%. Switch to mean-reversion entry signals. Tighter stops.
State 3: Crisis
σ > 30% · VIX contango → backwardation
Position sizing reduced to 20% of nominal. All new entries suspended. Existing positions tightened to 1× ATR.
Market Data → Rolling Features (σ, ACF, VIX slope) → HMM Forward Algorithm
→ Posterior State Probabilities → Regime Classification → Adapt Parameters
HMM Regime State Transitions
Figure 2: Posterior regime probabilities, Jan–May 2020.
The Response: Automated Adaptation
On February 24, 2020 — 48 hours after the initial sell-off — the HMM posterior probability for the Crisis state crossed the 0.7 threshold. The system automatically:
- Reduced position sizing from 100% to 20% of nominal allocation
- Suspended all new momentum-following entries
- Tightened stop-losses on existing positions to 1× ATR (from 2.5× ATR)
- Logged the regime transition with full state probabilities for audit
No human intervention was required. The system executed the adaptation within 200ms of the regime detection signal.
Result
The adaptive system preserved 75.9% of capital that would have been lost under static parameters. The regime detection lag of 48 hours — while not instantaneous — captured 91% of the protective benefit compared to a theoretical perfect-foresight oracle.
Build Regime-Aware Systems
Talk to a lead architect today about making your infrastructure survive the next regime break.
Schedule Technical Assessment