Autocorrelation
Quick Reference
| Property | Value |
|---|---|
| Dimension | stability |
| Category | statistics |
| Version | v1.0 |
| Output Column | autocorrelation |
Autocorrelation of any input signal at specified lag 鈥?measures signal persistence
Formula
autocorr(signal, lag)
CDM Inputs
| Column | CDM Table | Description |
|---|---|---|
signal | cdm_* | CDM source table |
Parameters
| Parameter | Type | Default | Description |
|---|---|---|---|
lag | integer | 2 | Lag for autocorrelation in milliseconds |
Output
Column: autocorrelation
Correlation of signal with its lagged copy
Market Intuition & Trading Rationale
Autocorrelation measures a signal's self-predictability: corr(signal[t], signal[t-lag]). It is the most important stability diagnostic 鈥?every signal in a feature set has autocorrelation measured to assess whether the signal is persistent (trending), mean-reverting (oscillating), or efficient (independent). High positive autocorrelation (> 0.3) means the signal is sticky 鈥?today's value strongly predicts tomorrow's. High negative autocorrelation (< -0.3) means the signal oscillates. Near-zero means each observation is independent.
The lag parameter determines the prediction horizon. Lag-1 autocorrelation captures tick-by-tick persistence 鈥?is this update similar to the last? Lag-10 captures medium-term persistence. The default lag of 2 captures short-term structure without being dominated by bid-ask bounce (which creates spurious negative lag-1 autocorrelation in price-based signals).
Usage Cases
- Signal quality filter: Signals with |autocorrelation| > 0.7 are highly redundant 鈥?they add little new information beyond what a simple autoregressive model would capture. Consider removing or downweighting in ML models.
- Strategy regime selection: Positive autocorrelation 鈫?trend-following strategies. Negative autocorrelation 鈫?mean-reversion strategies. Near-zero 鈫?either strategy works; prefer the one with higher SNR.
- stability dimension: Used in virtually every feature set as the stability diagnostic. Each pack measures autocorrelation of its primary signal to determine whether the signal's behavior matches the pack's strategy pattern.
YAML Definition
name: autocorrelation
description: Autocorrelation of any input signal at specified lag 鈥?measures signal
persistence
category: statistics
dimension: stability
version: v0.9.0 (Beta)
required_inputs:
- signal
output_column: autocorrelation
output_description: Correlation of signal with its lagged copy
tags:
- stability
- autocorrelation
- statistics
parameters:
lag:
type: integer
description: Lag for autocorrelation in milliseconds
required: false
default: 2
formula: autocorr(signal, lag)