Mean Reversion · When deviations repair
Distinguish temporary price pressure from permanent information changes and test a reversion hypothesis.
Observation
A rebound after a rapid decline may repair a liquidity shock, or merely interrupt a downtrend. Define what price should revert to: its past mean, similar assets, or redeemable value. These are three different economic anchors.
Mechanism
Urgent sellers accept discounts. Liquidity providers bear inventory and further-decline risk and may later receive compensation. Permanent fundamental information shifts the center, making the old mean an unreliable fair price. A mechanism needs a reason for demand to recover, not merely “it fell a lot.”
Hypothesis
The teaching experiment estimates short-term volatility from the past 20 daily returns. After an abnormally negative previous day, if the spread remains below a predeclared cap, buy at the next time and hold for at most two days. Fix invalidation and loss budgets. Determine thresholds in training, select once in validation, and freeze in testing.
Data
Use then-visible returns, spreads, volume, and halt/quote-interruption flags. Close-only data omits the inability to execute when execution matters most. Keep major events and delistings; do not delete “abnormal declines” afterwards.
Signal
Standardize deviations using means and standard deviations from past windows, never the full sample. With zero standard deviation, inadequate history, or stale quotes, output “No trade” and record the reason. A volatility-regime filter must be computable before entry.
Backtest
Compare exit horizons, cost stress, and an unfiltered baseline. Record unfilled orders and stop overshoots. After a stop triggers, use the next executable price rather than guaranteeing an exact threshold fill. Report correlated losses from repeated bottom-fishing and cap total risk of overlapping signals.
Falsification
Stress permanent price jumps, wider spreads, and absent recovery buying. If the rule works only after deleting crash days, tail risk contradicts the mechanism. If longer holding merely hides unrealized losses, it is not successful reversion.
Exercise and project acceptance
Prepare two teaching paths: a fall followed by return to the original center, and a fall followed by remaining at a new center. Apply identical entry/exit rules and daily accounting. Explain which observable condition could distinguish them and which cannot be distinguished beforehand. Include a loss budget, holding limit, shutdown trigger, and rejection record. Without an anchor mechanism, return to Observation.
Return to eight-category research. See Advanced A for statistical uncertainty.