Here is the follow-up to my previous post which shows out of sample results over the last week, 19th to 23rd of October.
Again a somewhat mixed bag, but with one huge caveat: this is the latest model of TabPFN over a smaller train data set. In my previous post the train data set was 17 week's worth of data and the TabPFN model used was version 3.0. In between then and now I updated to version 3.5, which in general is a bigger model, and I found that I was pushing the memory limits of my GPU with that much training data and then predicting on a following week. As a result, I found I had to reduce the amount of training data to be able to use my current framework, which currently means training on 1 week's worth of data to predict on the following week, using one model for all types of price behaviour.
This has two profound implications for my use of TabPFN:
- I will need to be much more parsimonious with my feature set to reduce the memory footprint so that I can include a longer history of training data
- I will probably need to train a series of more specialised models, e.g. one model for cycling price data, one for trending price data etc.
No comments:
Post a Comment