Description
Model the time between successive events when short gaps follow short gaps and activity follows a daily cycle.
Events such as trades, clicks or messages arrive at irregular times, and the gaps cluster: bursts of quick actions alternate with lulls, on top of a daily rhythm. Engle and Russell applied the GARCH idea to these durations. First the deterministic time-of-day pattern is estimated and divided out. The adjusted duration is then modelled as its conditional expected value times a positive random innovation with mean one, where the expected value is a weighted recursion on past durations and past expectations. The innovation can be exponential, Weibull or generalised gamma, which implies different hazard shapes. Persistence (the sum of the recursion weights) shows how long bursts last. Hazard and AFT models treat durations as independent, so this model fills the gap when inter-event times depend on recent activity.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Regress log durations on smooth functions of clock time, plus indicators for special periods such as opening and closing hours, and divide each duration by its fitted daily pattern.
- Remove or handle zero durations (simultaneous events), for example by merging them.
- Specify an ACD(r, s) recursion for the expected adjusted duration and choose an innovation distribution (exponential, Weibull or generalised gamma).
- Estimate by conditional maximum likelihood.
- Check that standardised innovations and their squares show no remaining Ljung-Box autocorrelation and match the assumed mean and spread.
- Interpret persistence, expected duration paths and the hazard shape, and use the model to forecast the next gap.
Sources
- Analysis of Financial Time Series (Tsay, 2002)
Also known as
- ACD model
- EACD
- WACD
- GACD
- Duration clustering model
- Autoregressive conditional duration (ACD) model for inter-event times
Categories
- Analytics & Behaviour
- Statistics & Modelling