Lookahead bias ?

[polizzi4757]

First of all I want to thank you for sharing your package. I think you have done a great job. However, if your goal is to fit HMM to financial data, I have to tell you that the HMM calculations might have a lookahead bias. To show you what I mean, I attach two scripts:

• "daily HMM calculation" where the model is recalculated for each day in the sample and only the state of the day under recalculation is considered and is not changed in subsequent recalculations (i.e. signals a calculated only with information available at the moment of calculation). Since this calculation needs time, I attached an rds file with the output of this script.
• "Compare HMM signals" where the daily HMM calculation is compared graphically with the model calculated once, where each signal uses all the available information in the sample, even future information, if I understand correctly.

It would be nice to have a feedback from you. Thank you Giuseppe Polizzi

HMM lookahead bias.zip

[loelschlaeger]

I was able to reproduce the bias with your code. I strongly believe that this has nothing to do with the implementation, but rather with the underlying model and especially the applied Viterbi algorithm for global decoding. So when fitting an HMM to a financial time series and subsequently applying the Viterbi decoding, we can only make probabilistic statements about future states and future (log-) returns, cf. https://doi.org/10.1177/1471082X211034048. Recalculating specific dates definitely compromises the global decoding, so your result is very much expected. I guess local decoding is required in this case, and we will put it on our "TODO" list for future package developments. Thanks for pointing this out, Giuseppe.

[polizzi4757]

Thank you for your timely response and thank you again for your great work.