Skip to reversion research
Mean Reversion StrategyReference / Excursion / Evidence

Causal z-scores and the overshoot problem

Define a prior-window reference, measure absolute recovery and avoid calling every correctly directed move a successful reversion.

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.

Specify what the current close can know

A rolling z-score starts with an observation contract. Name the instrument, adjustment policy, bar calendar and exact point at which a closing value becomes available. At index t, the reference window should contain the declared earlier observations, not the later close you hope to predict. Our lab uses the preceding window and a population standard deviation for that window. Another convention is possible, but it must be named before comparing results. A vendor correction received tomorrow was not available at today's decision. Save the original source and revision clock when those distinctions matter.

Keep the destination unchanged for the question

Suppose a prior-window mean is 100 and the current close is 104. The initial absolute distance is 4. A next close of 102 has distance 2 and counts as closer. A next close of 95 moves downward but leaves distance 5, so it does not count as recovery. A next close of 96 has the same distance 4 and fails a strictly-closer definition too. Replacing the mean with tomorrow's rolling value changes the question. You may study a moving destination, but label it separately rather than changing the definition halfway through a record.

Separate scale from probability

Dividing a displacement by an earlier-window standard deviation makes unlike price magnitudes easier to describe. It does not make the statistic normally distributed or supply a calibrated tail probability. A small scale can turn modest price movement into a large z-score; a constant window has zero scale and no defined z-score. Report those exclusions instead of assigning a convenient zero. Skew, outliers, dependence and estimation uncertainty all affect interpretation. A two-standard-deviation threshold is a rule chosen by the researcher, not evidence that the next observation is overwhelmingly likely to reverse.

Choose the denominator before looking at outcomes

The lab baseline includes eligible next-close observations with a non-zero prior-window scale. The stretch subset applies the declared absolute z-score threshold. Keep total input observations, eligible windows, zero-variance exclusions and stretch count distinct. Do not compare ten selected stretches against all raw input bars while concealing the warmup and last-bar exclusions. Adjacent stretches may describe the same extended excursion. They are not automatically ten independent opportunities, and a descriptive percentage should not be dressed up as a strategy win rate. Record missing bars and explain whether the window counts observations or elapsed time.

Do not promote the directional move into a fill

The lab also shows a signed next-close price change opposite to the displacement direction. That is a diagnostic quantity, not the return of a hypothetical short or long account. A trade would need a known decision time, attainable entry, exit instruction, costs and position size. An overshoot can have a positive directional move and still fail the distance test. Conversely, a closer next close need not produce a profitable trade after friction. Keeping both measures visible is useful precisely because they answer different questions rather than because they can be combined into a persuasive score.

Freeze an experiment you can reject

Start with one window, threshold, recovery definition and reserved period. Save the complete event list, not just pleasing examples. Describe the adverse excursion and unresolved-event handling before extending the horizon. If later exploration chooses a new scale or threshold, preserve the original result and mark the inspected period as used. The synthetic sample demonstrates software behaviour only. A clean calculator test is not an independent market experiment. Before asking whether the idea trades well, confirm that the reference is causal, the denominator is stable and the result survives the exact definition you initially wrote down.

A small rejection fixture

Construct a prior window of 98, 99, 100, 101 and 102. Its mean is 100. Append current 104, then try next closes 102, 96 and 95 separately. Record distance before and after, not just the sign of the move. The first is closer, the second ties and the third is farther. Run the same fixture after changing only later observations; the original prior-window mean must stay unchanged. That is a software causality check, not market evidence.

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

From one next close to an inspected horizon

The overshoot fixture preserves the original distinction: initial 104 and reference 100 become 95, so directional movement occurs while distance grows from four to five. Later observations can get closer, but their success cannot relabel that first next-close outcome. The new lens records each endpoint separately. Its authored rows carry no actual price source and no probability of future recovery. A genuine multistep study needs its complete event population and label-end ownership rather than a convenient extension after inspecting favorable later prices.