Description
Find whether one metric genuinely leads another, and by how many periods, by filtering out each series' own autocorrelation before cross-correlating, then fitting a dynamic regression.
Cross-correlating two trending or autocorrelated series produces large, smooth and often spurious peaks: two independent random walks routinely appear related. Box and Newbold showed a celebrated six-month lead of car production over a share index vanished once the series were properly modelled. Prewhitening removes this trap. Fit an ARMA model to the (differenced) input series, apply the same filter to both input and output, and cross-correlate the filtered series, whose significance can be judged against plus or minus two over root N. Significant lags reveal the delay and shape of the output's response, which guide the orders of a transfer-function model: the output as a lagged, possibly decaying response to the input, plus ARMA noise. The fitted model is checked and can forecast using the input as a leading indicator. Feedback from output to input invalidates the method.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Siloed
- Stage
- Post-launch, Strategy
- Data
- Quantitative
- Effort
- Days
Steps
- Plot both series and difference or transform them until they look stationary.
- Fit an ARMA model to the input series and keep its filter.
- Apply the same filter to the input (leaving white noise) and to the output.
- Cross-correlate the two filtered series and flag lags outside plus or minus two over root N.
- Read the delay and the shape of the response from the significant lags and choose the transfer-function orders.
- Fit the transfer-function model with an ARMA noise term; check that residuals are white and uncorrelated with the prewhitened input.
- Use the model to forecast, treating the input as a leading indicator only where it does not itself need forecasting and there is no feedback.
Sources
- The Analysis of Time Series: An Introduction (Chatfield, 2003)
Also known as
- Prewhitened cross-correlation function
- Transfer function model
- Box-Jenkins transfer function
- Impulse response estimation
- Dynamic regression
- Leading indicator analysis
- Cross-correlation with prewhitening and Box-Jenkins transfer-function (dynamic regression) identification of lead-lag relationships
Categories
- Analytics & Behaviour
- Statistics & Modelling