# Autocorrelation — Self-Correlation at Fixed Lag
Online autocorrelation coefficient at a configurable lag.
"Is this signal trending or mean-reverting?"
| Update cost | ~12 cycles |
| Memory | `8×LAG + 32` bytes |
| Types | `AutocorrelationF64`, `AutocorrelationI64` |
| Priming | After `LAG + 2` samples |
| Output | `correlation()` in [-1, 1], `covariance()` — both `Option` |
| Error handling | Returns `Result<_, DataError>` on NaN/Inf input |
## What It Does
Computes the Pearson correlation between a signal and a lagged copy
of itself: r(k) = Cov(x_t, x_{t-k}) / Var(x).
- **r > 0**: trending / persistent — values tend to continue in the same direction
- **r < 0**: mean-reverting / anti-persistent — values tend to reverse
- **r ≈ 0**: no linear dependence at this lag (random walk)
## When to Use It
**Use Autocorrelation when:**
- You want to classify a signal as trending vs mean-reverting
- You want to detect periodicity at a known lag
- You need to monitor stationarity (stable autocorrelation = stationary)
**Don't use Autocorrelation when:**
- You need correlation between TWO different signals → use [CrossCorrelation](cross-correlation.md)
- You need to detect which lag has the strongest effect → use CrossCorrelation with multiple lags
- You need nonlinear dependence → autocorrelation only captures linear relationships
## How It Works
```
State:
buffer[LAG] — circular buffer of previous values
mean — running mean (Welford)
m2 — running variance accumulator
cross_m — cross-moment accumulator
On each sample x (after buffer is full):
x_lagged = buffer[oldest]
// Update Welford mean/variance
delta = x - mean
mean += delta / count
m2 += delta × (x - mean)
// Update cross-moment (approximate: uses current mean for lagged value)
cross_m += (x - mean) × (x_lagged - mean)
// Store in buffer
buffer[head] = x
head = (head + 1) % LAG
Autocorrelation r(k) = cross_m / m2
(1/N normalization cancels in numerator and denominator)
```
The cross-moment uses the current running mean for both x(t) and
x(t-LAG). This introduces O(1/n) error that is negligible for
streaming (n >> 100). Exact computation would require separate
means for the lagged and current windows.
## Configuration
```rust
use nexus_stats::signal::AutocorrelationF64;
// Lag-1 autocorrelation (is the signal trending?)
let mut ac = AutocorrelationF64::builder().lag(1).build().unwrap();
for sample in data {
ac.update(sample).unwrap();
}
if let Some(r) = ac.correlation() {
println!("lag-1 autocorrelation: {r:.3}");
}
```
`lag` determines the buffer size and is set at construction time
via the builder.
## Examples by Domain
### Trading — Regime Detection
```rust
let mut ac = AutocorrelationF64::builder().lag(1).build().unwrap();
// Feed returns:
ac.update(price_return).unwrap();
if let Some(r) = ac.correlation() {
if r > 0.3 {
// Trending regime — momentum strategies work
} else if r < -0.3 {
// Mean-reverting — fade the moves
} else {
// Random walk — no edge from autocorrelation
}
}
```
### Monitoring — Stale Feed Detection
```rust
// High autocorrelation in quote changes = not updating
let mut ac = AutocorrelationF64::builder().lag(1).build().unwrap();
// Feed quote deltas (should be ~random if feed is live):
ac.update(new_quote - last_quote).unwrap();
if let Some(r) = ac.correlation() {
if r > 0.8 {
alert("feed may be stale — changes are highly autocorrelated");
}
}
```
### Networking — Jitter Periodicity
```rust
// Detect periodic jitter at known interval
let mut ac = AutocorrelationF64::builder().lag(100).build().unwrap();
// Feed inter-packet intervals:
ac.update(interval_ms).unwrap();
if let Some(r) = ac.correlation() {
if r > 0.5 {
// Strong periodicity at lag 100 — possible scheduling artifact
}
}
```
## Integer Variants
`AutocorrelationI64` accepts integer samples,
convert to f64 internally, and return f64 correlation values. Use these
when your input is naturally integer (tick counts, queue depths) and
you want to avoid the caller-side cast.
## Performance
| `AutocorrelationF64::update` (lag=1) | ~12 cycles |
| `AutocorrelationF64::update` (lag=10) | ~14 cycles |
| `correlation()` query | ~3 cycles |
The update cost is dominated by the Welford mean/variance update
(one division). Buffer management is a single array write + modular
increment. The correlation query is one division.