Skip to reversion research
Mean Reversion StrategyReference / Excursion / Evidence

Mean-reversion half-life: assumptions before estimates

Understand positive AR(1) decay, observation clocks, coefficient sensitivity and why half-life is not a trade deadline.

Original editorial revision: 9 October 2026. Expanded edition: 14 October 2026.

Educational diagnostics, not a signal service. A standardized price displacement is not a calibrated probability; invented examples and user-supplied observations are not verified market performance.

Write the conditional model in its own units

An illustrative AR(1) displacement process can be written x next = phi times x current plus an innovation with conditional mean zero. Here x is distance from a fixed equilibrium expressed with sign, not necessarily the raw price. To interpret decay, specify the equilibrium, treatment of intercept, observation interval and innovation assumptions. A positive coefficient below one causes expected displacement to shrink geometrically in this model. That mathematical fact does not establish that the chosen market follows the model. The public worksheet deliberately asks for phi rather than pretending to estimate it from an unseen dataset.

Derive the fractional half-life

Iterating the conditional expectation gives phi to the power k times the starting displacement. Setting that multiplier to one half yields h = ln(0.5) / ln(phi). For an assumed phi of 0.8, h is about 3.1063 observation intervals. With fixed 60-minute observations, the expectation half-life is about 186.377 minutes. This is a fractional model interval: the third observed close has not quite halved the expected displacement while the fourth has gone further. Rounding down and presenting three observations as a guaranteed arrival would change both the arithmetic and its meaning.

Do not confuse decay with first passage

Half-life describes an expected signed displacement. First passage describes a random path reaching a boundary. Innovations can move a path farther away, past the reference or across a stop before any desired arrival. The half-life is therefore not a median recovery time, deadline for a successful trade or stop-loss setting. A trade horizon also depends on adverse risk, entry availability, costs and the predeclared execution policy. Those objects need separate analysis. A fast conditional decay can coexist with difficult fills or unfavourable risk, and a slower decay does not establish that an idea should be rejected by itself.

Keep the observation clock honest

Changing from hourly observations to five-minute observations changes the coefficient being modelled. Multiplying an hourly half-life by five minutes does not create a validated five-minute model. Irregular trading calendars, missing observations and overnight gaps also complicate the mapping from index intervals to elapsed time. Our worksheet assumes a fixed interval and simply multiplies by it. In a real study, retain exchange calendar, timezone and availability metadata. A completed-session counter, continuous wall-clock horizon and count of qualifying bars are different clocks; never let a convenient label hide which one the experiment uses.

Inspect fit sensitivity without opening a tournament

The statsmodels autoregression documentation provides a way to inspect model specification, lags and fitting. Decide the fitting window and selection procedure before inspecting reserved performance. A coefficient near one creates a long and highly sensitive half-life: assumed phi 0.95 gives about 13.5134 intervals. Parameter uncertainty and a changing reference can dominate a tidy point estimate. Preserve the fit version and residual diagnostics. Repeatedly choosing the interval, lookback and lag that makes the shortest half-life is an exploration of many candidates, not independent confirmation of one predeclared hypothesis.

Respect the worksheet boundary

Phi zero, negative phi and phi at or above one are rejected. Negative coefficients can alternate signs and need a different interpretation; the worksheet is not a general autoregressive simulator. Its chart is invented coefficient arithmetic rather than an empirical distribution of market recoveries. Export the scenario alongside your research notes and state what is still missing: causal data, a justified equilibrium, a frozen fit, independent evaluation and executable costs. A formula passing software fixtures verifies implementation. It does not show that a named provider, asset or timeframe has an economically usable mean-reversion process.

A coefficient exercise with no market attached

Set the interval to sixty minutes and compare assumed phi 0.5 and 0.8. One has an expectation half-life of one interval and the other about 3.1063 intervals. Then enter phi 1 and verify rejection. Save the exports under an invented-scenario label. Write a separate paragraph explaining why a path can cross a stop even while its conditional expected displacement shrinks. No historical prices are needed to learn this distinction; no trading conclusion follows from the exercise.

Source documents

Documentation establishes the stated mechanism, not tested trading performance. Checked 9 October 2026; a later version can differ.

Continue the investigation

AR(1) expectation half-life worksheet

Retain conditional expectation ownership

The path workbook deliberately has no estimated coefficient. Its observed first hit can differ from any assumed expectation half-life, and a later rebound can reverse a favorable intermediate distance. Keep the actual horizon and reference fixed when describing that contrast. It would be incorrect to infer a passage distribution or choose a profitable stop from these seven authored cases. A real coefficient needs its actual fit scope, source and reserved evaluation; a real strategy additionally needs native exits and accounting. Adding the full path diagnostic does not replace the original conditional formula or weaken its limitations.