A0794
Title: ARDL models with high frequency intervals and dynamic sampling: A case study from crypto liquidity
Authors: Daniel Finnan - Conservatoire national des arts et metiers (CNAM) (France) [presenting]
Abstract: Cointegration analysis and ARDL models are typically applied to quarterly or monthly data, with inherently less noise and variance than more granular intervals. However, the bounds testing procedure at the heart of ARDL analysis can equally be applied to higher frequency data if the model being explored has a strong theoretical grounding. Furthermore, with a greater sample size, techniques employing dynamic sub-samples can be employed to capture changes in the cointegrated relationship, using rolling windows and recursive sampling. This is akin to testing for structural breaks, but places the emphasis on the F- and t-statistics in the bounds testing procedure, rather than traditional tests for stability such as CUSUM. It also has the advantage of enabling the econometrician to build up a picture of changes in the speed of adjustment and long-run relationship over time. When applied to time series data using a variety of frequencies, it is possible to establish results giving both an overall view and a much more detailed, fine-grained perspective of changes to a relationship. An illustration of the approach is presented in the form of liquidity flows on cryptocurrency exchanges, using both daily and hourly intervals, over a year, month-by-month, and using dynamic sampling.
