I’m currently working on a xgboost model to predict the price change above or below a given percentage between a candle’s open price and the next candle’s close price. I use a wide range of features, including market sentiment, unemployment rate, inflation, s&p ohlc data, as well as calculated technical indicators for the coins at hand. I have seen a recent article on achieving performance vastly superior to hodl on ETH and BTC with boosting models in particular (https://arxiv.org/pdf/2311.14759.pdf), but I’ve grown skeptical to whether these numbers could be achieved in practice. The question is to the people who have implemented similar methods for work or for their own projects — what are the considerations in terms of using boosting for time series analysis and specifically algo trading? Are boosting models even used in the industry for such tasks? Thanks in advance to whoever sheds some more light on this matter, I’d be very grateful!

  • $\begingroup$ I have not read this research paper. But generally when stupendous results are achieved by academics on past (in-sample) data they tend to fall apart on out-of-sample (oos) data that the model has not encountered before. Which kills practical application of the model (but is ok if the goal is to get a paper published). $\endgroup$
    – nbbo2
    Feb 15 at 14:43
  • $\begingroup$ @nbbo2 Thanks for the reply! That's what i thought too. Do you know whether GBM methods are even used for such tasks (trend prediction/signal generation) in quant hedge funds or other similar organizations? i've heard that people mostly prefer simpler linear models, but to what extent that is? Thanks! $\endgroup$ Feb 15 at 19:44


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.