Causal Structure, Not Just Parameters: Why Regime Switching Matters for Options

The misconception

Different market regimes are not different parameter values for the same model. They are different causal structures. A trending market and a range-bound market are driven by fundamentally different forces: trend-following hedging costs versus bidirectional risk premia wrestling arbitrage. Stuffing a regime indicator into one model asks it to implicitly classify which force dominates, and that implicit guess is harder to get right than an explicit regime classification you design yourself.

There is a common belief that if you simply feed a regime indicator into a model as an input variable, the model will automatically adjust its judgments for different market conditions. This idea overlooks something fundamental: the forces driving options pricing are not the same logic across regimes. You cannot make one universal model work just by tuning parameters.

Two regimes, two causal structures

Consider a stock in a sustained trend, up or down. Dealer hedging actions follow a clear pattern: price moves in one direction, hedge positions adjust in the same direction, no need to frequently reverse. Hedging costs stay low. That low cost shows up in implied volatility as compression.

Now switch to a range-bound market. Everything reverses. Liquidity providers must charge an additional risk premium for the uncertainty of having to reposition in either direction at any time. And once price deviates from a range, arbitrage flows arrive to pull it back. That tug-of-war itself pushes short-term volatility higher.

These two environments are not a matter of the same variable being a bit larger or smaller. The forces supporting the price are structurally different things. One is a trend-following hedging cost structure. The other is a bidirectional risk premium wrestling against arbitrage forces. Trying to capture both with a single formula whose parameters you tune is asking that formula to implicitly guess which force is dominant, and that implicit guess is often harder to get right than an explicit regime classification you design yourself.

Explicit classification means you choose which indicators to use, set appropriate thresholds, and partition market conditions into discrete categories. The process is transparent: you can articulate exactly why a given day belongs to a given regime. An implicit classification buried inside a unified model gives you no such transparency, and when it misclassifies, you cannot diagnose why.

The academic foundations

Hamilton (1989) introduced Markov-switching models to capture structural breaks in economic and financial time series, establishing the regime-switching framework: market states are discrete and identifiable, not points on a continuous smooth curve. The model estimates regime-specific parameters and a transition probability matrix, allowing expansions and recessions to emerge as distinct probabilistic states rather than gradations of a single process.

Building on this, Duan, Popova and Ritchken (2002) showed that Black-Scholes is a special case of a regime-switching framework: the single-regime case with no feedback from returns to volatility. A unified model covering multiple regimes is, in effect, being asked to perform implicit regime classification internally. That internal classification is often harder to handle than an explicit one you design yourself. (The paper was published in Quantitative Finance in 2002; earlier working-paper versions circulated from 1999.)

The evidence — and the disagreement

Whether regime-switching models are necessarily more accurate than constant-volatility models is not settled in the literature. Two studies, different markets, different conclusions.

Kilander (2007), in a master's thesis at the Stockholm School of Economics, calibrated a two-volatility-state regime-switching model to OMXS30 index options. The result: pricing errors were significantly lower than those produced by a standard Black-Scholes model. Multiple volatility states improved the fit to observed call-option prices.

But Kalovwe (2023), published in Cogent Economics & Finance, found the opposite, with an important nuance. Using Russell 2000 and Facebook (Meta) option data, the study compared Black-Scholes, a regime-switching (RS) model, and a regime-switching GARCH (RS-GARCH) model. For short-dated options (25-day contracts), Black-Scholes produced the lowest root-mean-square error. For long-dated options (258-day contracts), RS-GARCH outperformed both Black-Scholes and the plain RS model.

This is not a contradiction that weakens the case for regime-switching. It strengthens it, but for a different reason than "it's more accurate." The inconsistency tells us that regime-switching's value is not that it is universally more precise. Its value is that it acknowledges a fact about markets: structure changes. Which model is closer to reality depends on the asset, the period, and the contract maturity. You cannot assert one is always better.

A four-quadrant framework for options

The academic literature tells us regimes exist and matter. The practical question is how to classify them operationally. The framework below uses two observable dimensions, adapting the original design by Voltima_quant, whose work on regime-conditional strategy selection inspired this translation to options. The original framework is theirs; what follows is a domain translation and academic supplement.

Two dimensions define the quadrants:

Volatility level, measured by the rolling percentile of realized volatility (RV percentile). Where does current realized vol sit relative to its own recent history?

Trend strength, measured by the Hurst exponent or the more common ADX indicator. A Hurst exponent above 0.5 indicates persistence (trending); below 0.5 indicates mean-reversion tendency. ADX above 25 generally signals a clear trend; below 20 suggests range-bound conditions. This classification logic is consistent with Hamilton's (1989) framework: market states are discrete and identifiable, not points on a continuous curve.

Crossing these two dimensions produces four quadrants:

Quadrant 1: Low vol, low trend — range-bound grind

Implied volatility is thin. Narrow premium-selling structures collect limited credit. The expected value of the strategy is compressed from the start: you are taking on undefined risk for a small premium in a market that is already calm.

Quadrant 2: Low vol, high trend — slow grind

Implied volatility is low, but price is moving directionally. Buyers are relatively well-compensated: directional debit structures make more sense than selling premium. In this quadrant, the premium a seller collects often cannot compensate for the directional risk they assume.

Quadrant 3: High vol, low trend — panic chop

Implied volatility is elevated but price has no clear direction. For premium selling, this is the ideal environment: rich premiums without trending risk. Lerman (2018), in a CME Group educational article, proposed a practical framework: use the historical percentile rank of implied volatility to time entries. The article suggests considering premium selling when IV sits above the 75th–90th percentile of its recent distribution, and considering premium buying only below the 10th percentile, rather than looking at absolute volatility levels, which are not comparable across assets.

Quadrant 4: High vol, high trend — crisis trend

This is the hardest quadrant to handle, and the most important. Implied volatility is spiking at the same time price is moving sharply in one direction. Dealer hedging liquidity dries up. Traditional premium-selling strategies in this quadrant tend to record amplified losses: the very structure that collects premium in calm markets is the one that gets punished when vol explodes and direction persists.

Lerman (2018) cited a real case study that illustrates this precisely. A client sold a large straddle position on S&P 500 futures when volatility was at 7.1%, the 2nd percentile over the past decade. Statistically, 98% of the time volatility was higher. Five trading days later, volatility jumped from the 2nd percentile to the 90th. Both the put and the call in the straddle exploded in premium terms. The position recorded a major loss in under a week.

This case demonstrates the core point: Quadrant 4 occurs least frequently in history, but when it arrives, its destructive power against premium-selling strategies far exceeds the other three quadrants combined.

Sample imbalance: the real risk management problem

The high-vol, high-trend regime is the rarest in historical data, and the most consequential. This is a classic sample-imbalance problem. If you try to fit a dedicated implied volatility surface or Vega risk model to this quadrant alone, you are working with very few observations. Overfitting is almost guaranteed.

The more defensible approach is to treat this scarce regime with limited stress-scenario analysis rather than a fully fitted model, and to deliberately reduce strategy complexity when you detect it. Prefer defined-risk debit spreads or protective put purchases over naked short positions or complex multi-leg credit structures. You are not trying to extract alpha from the rarest, most dangerous regime. You are trying to survive it.

Where this meets our work

Regimes are structurally different, not parametrically different. That principle shapes how we build at RiskHarvest: classify regimes explicitly and apply the methodology suited to each, rather than constructing a single model that implicitly handles all market conditions. Three specific mechanisms reflect this:

Vol-targeted risk sizing scales exposure to realized volatility, not a fixed allocation. The same target return requires different gross exposure in a low-vol regime versus a high-vol regime. This is regime-conditional by design, not a parameter tuned inside a static model.

Portfolio-level rebalance triggers fire based on capital-efficiency conditions, not a fixed calendar schedule. The decision to act depends on the state of the portfolio relative to its environment, not on the passage of time. This is explicit regime-aware logic.

Premia decomposition breaks a holding's return into the underlying risk premia it exposes you to: equity, term, volatility, carry. Two assets that look different may be carrying the same premia, and a regime shift that hurts one premia hurts both. Knowing the structure, not just the parameters, is what makes diversification real.

The one thing

Different market regimes are not different parameter values for the same model. They are different causal structures. Treat them explicitly. The academic evidence does not say regime-switching is always more accurate. It says markets change structure, and pretending otherwise is the real risk.

This article adapts the original regime-conditional framework by Voltima_quant, whose work on regime-based strategy selection inspired this domain translation to options and academic supplementation. The original framework is theirs; errors in translation are ours.

Sources

• Hamilton, J. D. (1989). "A New Approach to the Economic Analysis of Nonstationary Time Series and the Business Cycle." Econometrica, 57(2), 357–384. JSTOR
• Duan, J.-C., Popova, I., & Ritchken, P. (2002). "Option pricing under regime switching." Quantitative Finance, 2(2), 116–132. Taylor & Francis
• Kilander, M. (2007). "Calibrating an option pricing model under regime-switching volatility." Master's thesis, Stockholm School of Economics. SSE Archive
• Kalovwe, S. K. (2023). "On regime-switching European option pricing." Cogent Economics & Finance, 11(1). Taylor & Francis
• Lerman, G. (2018). "Volatility in the Crosshairs: Aligning Volatility and Strategies." CME Group Educational Article. CME Group