ACD and Long-Memory Models for Duration Forecasting

Summary

The document discusses autoregressive conditional duration models for transaction times and related nonnegative market data. A practitioner reports that, in tests forecasting bid–ask spreads, the error distribution mattered: Burr errors performed better than exponential, Weibull, and generalized gamma alternatives. The practitioner also reports stronger out-of-sample density forecast accuracy from long-memory variants, including FIACD and LMACD, than from other multiplicative error models.

The response suggests testing long-memory specifications when a basic ACD(1,1) has weak out-of-sample performance, and says multiplicative error models may be preferable to ARMA models for forecasting tails. These observations come from one reported set of comparisons, focused on spread forecasts rather than transaction-duration forecasts in the question. They do not establish that the same ranking will hold across markets, samples, forecast horizons, or evaluation criteria. The answer mentions planned estimation code but does not provide it in the document.

Key ideas

• The reported forecast comparisons found that the assumed error distribution materially affected model performance.
• Burr errors performed best among the distributions tested in the practitioner’s spread forecasting exercise.
• FIACD and LMACD long-memory models reportedly improved out-of-sample density forecasts relative to other tested multiplicative error models.
• The practitioner reports that multiplicative error models may be preferable to ARMA models for tail forecasts.
• The results concern one empirical exercise and may not transfer to every duration dataset.

Tags

# Use of ACD to model transaction durations


# Use of ACD to model transaction durations












I am using a simple ACD (autoregressive conditional duration) model with expoential or Burr distributed residuals and 1 lag, i.e. ACD(1,1).

I am modelling durations for transactions data on a 'medium' liquidity market.

I obtain a correct fit, but not very good out-of-sample results. Does anyone have experience with handling this model on real data? and using it for out-of-sample prediction.

I am interested by practitioner's ressource: code, tutorial, example, notebooks, etc. or direct feedback.

## Answer by Malick (score 1)

https://quant.stackexchange.com/a/31931

I have tested several Multiplicative Errors Models, including the standard ACD model. I used these models to forecast the bid ask spread ( non-negative data) and I obtain interesting results. As a main information I have found that the choice of the error term distribution is pivotal with the Burr distribution being by far the best (tested: Exp, Weibull, Generalized Gamma). Also I have found that long memory models (FIACD by Jasiak (1998), LMACD by Karanasos (2003)) have better out-of-sample density forecast accuracy than other MEM models. So I recommend you to test these latter models. Finally I have compared those models with ARMA ones and found that MEM models should be preferable for tails forecasts.

I plan to release freely the code (in Ox) that I made to estimate those models, I still need to do a minimal help documentation. I'll edit this answer when I'll do it.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.