TimeMixers as a new time series tool

 

Many trend-followers are time mixers. What has been old for some is now something new. A new technique in machine learning time series forecasting that is being given more attention is a concept called time mixing. The idea of time mixing is simple. Decompose any time series into a set of different time scales. For example, there can be short-term behavior, longer-term, seasonal and very long-term behavior in prices. This breakdown is certainly true for almost all commodities. If we can take the multi-scale data, we can break the problem into two parts or blocks:  the past-decomposable-Mixing block PDM where the information at different scales is learned by the model. From the PDM, the model is sent to the Future-Multipredictor-MIxing (FMM) block which will combine the scaled information into a forecast. This type of model may not do well for very short-term predictions, but longer-term forecasts can be improved by this technique. Given this is a ML technique, the value is found in improved forecasting. This is not a structural model of how the market behaves.

What do many trend-followers do? They will often use trends for different scales or look-backs to blend into a single signal. The weighting of these signals can be a very simple, like an equal-weighted scale, or it can be optimized based on preferences or on past forecasting success. This is not the sample as a formal time mixer model, but the idea behind this approach is the same. Use different time scales to capture the different behaviors in asset time series.