Mean reversion is a hypothesis that a defined price or indicator deviation tends to move back toward a chosen reference. The reference might be a rolling average, a prior range or a statistically estimated value. The phrase does not explain why a return should occur, how long it could take, or what happens if the deviation grows. Before testing, define each of those pieces and include the possibility that the idea simply fails.
Turn the phrase into an observable rule
Choose one instrument and price series. State whether it is spot or a particular perpetual, the bar interval, timezone, data source and period. Define the reference using only information available at the decision time. For a rolling average with window n, specify n and whether you use closes, typical prices or another input. For a deviation trigger, state its units: a percentage, a fixed price difference, a standardized residual or an indicator threshold are not interchangeable.
Then write the sequence. One example for research only: after a completed bar closes a predeclared distance below a rolling average, record a hypothetical long signal; use a later bar under a stated entry convention; exit when price crosses the average, reaches a maximum holding time, or violates a predeclared loss threshold. Explain what happens if several exit conditions could occur within the same bar. A rule that says “buy oversold” still needs to name the measurement, threshold, timing and exit.
The critical counterexample is persistence
A deviation can widen instead of closing. A market may continue moving in one direction because new information, changing liquidity or a market regime differs from the conditions used to select the threshold. A band touched in a chart is not evidence that price must return. If the rule averages down, its exposure may grow precisely while the original assumption is being challenged; define a total exposure cap and a clear stop condition rather than treating a lower average entry as proof.
Test the rule on chronological intervals and against a simple baseline. Record how often the signal occurs, the full return path, adverse excursion, holding time, drawdown and the number of simultaneous or consecutive losses. Examine whether a few recoveries account for most apparent gains. A high percentage of small wins can coexist with infrequent, much larger losses. Net each result of fees, spread, slippage and any applicable funding; do not infer success from gross price movement.
What the crypto evidence says—and leaves open
Studies can use the same phrase while testing different series, horizons and statistical properties. Turatti, Mendes and Caldeira’s Bitcoin study analyzed 443 weekly price observations from October 2010 to January 2019 with a three-state volatility model. They reported mean aversion in the tested multi-period log returns. That result belongs to that sample and method; it is not a test of every coin or a present-day moving-average rule.
A 2026 Finance Research Letters paper by Kar and coauthors uses quantile-panel persistence tests on major cryptocurrencies and reports asymmetry: downside shocks were relatively transitory while upside shocks tended to persist. It also reports significant dependence across assets. That is a conditional return-dynamics result, not a promise that a countertrend entry will earn money. The 2026 study and the earlier Bitcoin paper ask different statistical questions, so their results should not be collapsed into a single “crypto always reverts” rule.
For a strategy test, state whether the target is a price level, a return, or the sign and size of a recent move; choose the asset, venue, sampling interval and evaluation horizon; then freeze the threshold, entry, exit and cost model. If evidence appears only in one tail, regime or short horizon, report that boundary and test it on later data. Statistical predictability is not the same as an executable edge after spread, fees, slippage and funding.
Separate a signal from an explanation
An indicator can mechanically describe a past move without explaining what caused it. If the rationale is “price stretched too far,” define “too far” before viewing the outcome and name an alternative explanation such as a sustained repricing. Compare nearby thresholds and rolling windows without searching until one happens to fit. Reserve a later period in advance, and mark it as used once it influences a choice. Do not present the strongest variation as though it had been the original idea.
For a perpetual position held over funding settlements, include the actual historical settlement observations with their source, timestamp and sign. A rate observed today is not a promise about later cash flow. Confirm contract size, collateral, margin and liquidation mechanics on the selected venue. The CFTC warns that leveraged virtual-currency futures carry substantial risks; margin is not a cap on possible loss.
What SuperQuant can demonstrate
SuperQuant includes a simple RSI threshold-reversal rule among its research examples. The implementation generates a directional signal from a threshold crossing on a previous closed hourly bar; that is a specific product rule, not a universal mean-reversion definition. Its bundled data are synthetic. The engine’s results help inspect the configured logic and ledger under its assumptions; they do not establish how any real market behaves or whether reversion will occur.
Before running an experiment, record the exact trigger and exit, maximum holding time, position-size method, cost model, benchmark, failure condition and date range in a trading plan. Compare the assumptions against the broader trend-following guide, then use the backtest audit checklist. Mean-reversion research is educational and is not a signal or individualized investment advice.