Experimental regime research

What market state are we probably in — and is it likely to persist?

Regime Lab fits a transparent three-state Gaussian Hidden Markov Model to bundled SPY or QQQ history. The purpose is not to predict tomorrow's price. It is to estimate latent market-state probabilities, persistence, and expected duration so we can test whether regime awareness adds value to options-strategy decisions.

Most likely state

—

Probability

—

Expected duration

—

4-week persistence

—

Model setup

Price-only prototype · weekly observations · 3 latent states

Underlying

Training window

Features

Weekly return · 4-week momentum · 4-week realized volatility

Bundled data are adjusted-close log returns. State names are heuristic labels assigned after fitting; the latent states themselves are statistical clusters, not observed facts.

Training window sensitivity: The inferred current state can change when you change the training window. For example, SPY may classify as Calm/Bullish over 10 years but Choppy/Neutral over 5 years. This is evidence of model sensitivity to the choice of lookback period. Fit the model with different windows to understand the stability of the current classification before acting on it.

Current filtered state probabilities

—

HMM estimate

Fit the model to begin.

No live market feed is used; this page works from the bundled historical dataset.

Regime history

Reconstructed index from weekly returns

Calm / BullishChoppy / NeutralStress / Bearish

Regime-information research sequence

Stage 2 now complete

1Fixed strategy

Run the same option rules without knowing the regime. This remains the baseline.

2Perfect-state oracle

Give the strategy the true current state at each 30-DTE decision boundary. This measures the maximum value current-state information can add under the frozen policy.

3HMM-guided strategy

Use only filtered state probabilities and persistence available at the decision date, then compare with the oracle and fixed baselines.

Progress: HMM-ORACLE-001 found economically meaningful value from perfect regime knowledge when options were fairly priced. The next test is whether a real-time HMM can capture enough of that value without look-ahead.

HMM-ORACLE-001 · Perfect-state upper bound

5,000 paired paths · 36 × 30-DTE cycles

Covered-call test: 100 shares, 30-DTE 0.20-delta calls, realistic option fee/execution friction, no rolling. The hidden state is held fixed through each option contract and can transition only between contracts. The state-specific sell/hold rule was learned on a separate training sample, then frozen for the 5,000-path test.

Loading oracle experiment results…

Important model correction: a first weekly-switching prototype was rejected because an option could be priced in one regime and then carried into another before expiry. That created a surface/process mismatch and a false edge. The reported experiment removes that artifact by aligning each hidden state with the full 30-DTE contract.

State characteristics

Posterior-weighted estimates

Transition matrix

One-week transition probabilities

Persistence of current state

Probability of uninterrupted continuation

Interpretation

What the HMM does and does not say

Useful question: is the market currently behaving like a historically persistent state whose return/volatility distribution differs from the alternatives?

A high state probability is not a forecast that prices must rise or fall. HMMs infer hidden statistical states from observed features. The economically important test comes next: whether acting on those probabilities out of sample improves an option strategy after trading friction.

Stage 3: Walk-Forward HMM Strategy Test

Does real-time regime detection improve Wheel timing?

Finding: The walk-forward HMM-guided strategy captured 0% of the oracle's regime-awareness advantage across all IV premium scenarios. The model identifies persistent states but cannot extract an actionable signal for real-time strategy timing.

Stage 2 (Oracle): Perfect regime knowledge added +1.02 pp/yr at fair IV, confirming that regime awareness has theoretical value. Stage 3 (Real-Time HMM): A walk-forward HMM refit monthly using only prior data never accumulated enough confidence and persistence to change the fixed strategy rule. The decision thresholds (≥0.65 posterior probability + ≥20 weeks expected duration) were too stringent for practical signals.

IV Premium	Oracle Edge	Capture Ratio	Interpretation
0% (Fair)	+1.02 pp/yr	0%	Oracle has clear advantage; HMM missed it
5%	+0.44 pp/yr	0%	Moderate oracle edge; HMM silent
10%	−0.04 pp/yr	0%	Options rich enough; oracle neutral
15%	−0.38 pp/yr	0%	Oracle underperforms; HMM correctly defaults
20%	−0.83 pp/yr	0%	Option premium dominant; oracle disadvantaged

Protocol Status (per HMM-001): RED — Do not deploy as strategy-selection tool. The HMM remains valuable as a descriptive research feature for understanding how regimes correlate with strategy performance, but the real-time signal is insufficient to improve Wheel timing out of sample.

Implication: The regimes you see in Regime Lab are real statistical patterns. The oracle comparison shows what perfect timing could add. But in real trading, regime identification arrives too late or with too much uncertainty to act on. Use Regime Lab to understand your market environment and learn how strategies behaved historically in each state—not to time entries and exits.