Return Z-Score Reversion
Standardizes recent returns into a z-score; cleaner than price bands, but not a full stat-arb convergence engine.
This trades single-asset return shocks, not a richer cross-sectional or market-neutral statistical-arbitrage stack.
Implementation Scope
This trades single-asset return shocks, not a richer cross-sectional or market-neutral statistical-arbitrage stack.
The Intuition
The z-score mean reversion strategy is the most statistically clean version of mean reversion trading: it standardises the recent return relative to its own historical distribution, then trades when that standardised deviation exceeds a threshold. It is agnostic about price levels — it operates entirely in the space of standardised deviations.
The idea originates in statistics. A z-score of +2 means the observation is 2 standard deviations above the recent mean — an event that under normality would occur about 2.5% of the time. If returns are mean-reverting, a z-score of +2 is a bet that the extreme observation will be followed by a correction toward zero. The larger the absolute z-score, the stronger the signal.
This is closely related to the "convergence trade" logic underlying statistical arbitrage: you are not betting on a direction per se, but on the reversion of a statistical anomaly. Unlike price-level strategies (Bollinger Bands), the z-score approach works on returns, which are more naturally stationary than price levels.
Key assumptions: (1) Log returns are approximately i.i.d. within the rolling window — the window mean and variance are meaningful parameters. (2) Autocorrelation in returns is negative at short lags (mean reversion). (3) The threshold is appropriately set for the instrument's volatility regime. If returns are fat-tailed (as they typically are), the 1.5-sigma event is more common than normality predicts.
The strategy fails in trending markets and during regime changes. When a new information event causes a sustained shift in the price level, the z-score keeps rising rather than reverting. Position sizing is critical: because the strategy can be wrong during trending regimes for long stretches, naive full-position sizing leads to deep drawdowns. Many practitioners use Kelly-scaled or volatility-targeted position sizing alongside z-score signals.
The Math
Read this as a compact model summary: what the signal sees, what it ignores, and where fragility can creep in.
r(t) = log(Close(t) / Close(t-1))
mu(t) = mean(r[t-n : t])
sigma(t) = std(r[t-n : t])
z(t) = (r(t) - mu(t)) / sigma(t)
Signal(t) = +1 if z(t) < -threshold
= -1 if z(t) > threshold
= 0 otherwise
Parameters
| Parameter | Type | Default | Description |
|---|---|---|---|
| window | int | 20 | Rolling window for z-score calculation |
| threshold | float | 1.5 | Z-score magnitude to trigger a trade |
Source Code
def run(ticker: str, start: str, end: str, **params) -> dict:
window = int(params.get("window", 20))
threshold = float(params.get("threshold", 1.5))
df = fetch_ohlcv(ticker, start, end)
positions = _signal(df, window, threshold)
return run_backtest(df, positions, ticker=ticker, start=start, end=end,
strategy=METADATA["slug"], params={"window": window, "threshold": threshold})
Further Reading
- Chan, E. (2013). Algorithmic Trading, Ch. 2. Wiley.
- Vidyamurthy, G. (2004). Pairs Trading: Quantitative Methods and Analysis. Wiley.
- Avellaneda, M. & Lee, J. (2010). Statistical Arbitrage in the US Equities Market. Quantitative Finance, 10(7), 761–782.
When It Works / When It Fails
- Range-bound markets with approximately normal return distributions
- Instruments with stable mean and variance over the lookback
- Calm vol environments where z-score thresholds are meaningful
- Fat-tailed return distributions invalidate z-score thresholds
- Trending regimes — returns keep hitting the extreme threshold
- Non-stationary volatility periods distort rolling normalization