ARIMA and GARCH are old news for predicting volatility time series of asset returns. I am aware of papers that replace ARIMA and GARCH with machine learning algorithms to predict financial volatility more accurately, and so this question is a reference request for a survey of what's out there:

How do the performances of the following, and other, machine learning algorithms compare with one another, and GARCH, at forecasting volatility for horizons longer than just 1-day/step ahead?

  • random forest
  • support vector regression (SVR)
  • gradient boosting
  • K nearest neighbors, etc)

And are the machine learners also implemented in an autoregressive formula like GARCH (i.e. they use preceding historical volatility observations to estimate the current volatility)?

Also what do the research articles say regarding the theoretical reason for applying machine learning to volatility time series in particular?

  • $\begingroup$ Are you more interested in predicting the realised volatility of the time series over a period (which is backwards looking, we don't know it until a period is over), or the implied volatility (which is implied by the market in the price)? If you're after implied volatility, is VIX prediction sufficient (as a single number encapsulating 'vol') or are you interested in the whole surface? $\endgroup$
    – StackG
    Commented Aug 1, 2020 at 3:52
  • $\begingroup$ Just plain volatility of an asset's returns (the second statistical moment of a random variable), based on(the features being) historical estimates/realizations of itself, nothing about market proxies or derivative assets $\endgroup$
    – develarist
    Commented Aug 1, 2020 at 5:00

2 Answers 2


Just based on my understanding of the ML models themselves, I have a hard time believing KNN or RF are useful in anyway. They wouldn't be the first models I try and tend to just be ML models taught in class for who knows what reason honestly - maybe because they are easy to understand? From what I have read about ML in general (not in relation to time series), all of the ones you have listed have been outclassed by neural networks. Gradient boosting might be one that is still somewhat useful.

KNN looks to predict a value based on K observations that are most similar and then takes the average. Do you think tomorrow's volatility really is equal to the most similar days in your data set, even if the days are from 3 years ago? If so KNN, may be helpful.

RF is just a less good version of Gradient Boosting Trees. It predicts based on thresholds of features and as a result just partitions your data and then predicts based on the average of the average of many trees. So tomorrows volatility is equal to days when yesterday's vol is greater than x but less than y, snp moved by more than z but less than a, etc... Does that make sense? Maybe, but due to the partitioning nature of RF, it can never truly replicate any mathematical function. Meaning, if the true relationship between x and y is linear, a linear regression will always do better than a random forest.

  • $\begingroup$ your saying ANN has outclassed SVR, boosting and GARCH. Are they using autoregressive ANN? How many steps ahead can it forecast? references comparing machine learners would be good $\endgroup$
    – develarist
    Commented Aug 7, 2020 at 5:04
  • $\begingroup$ My understanding is that any model can be "autoregressive" if all you do is use a lagged variable as a predictor. Reason why traditional TS stats is it's own subject is that it has simplifying assumptions that allow you to calculate things like sampling distributions, unbiased estimators, confidence intervals, etc... These type of things you don't usually get from a ML model. For NNs, the flexibility is part of it's appeal - you can set up a NN to do a lot like predict multiple outputs, use multiple inputs, etc... $\endgroup$
    – confused
    Commented Aug 8, 2020 at 17:18
  • $\begingroup$ Unfortunately I don't have a good research paper that compares how different models work on TS data, I've been looking for one myself, but this is just based on what I've gathered reading different books and material on the stuff. And the fact that skills companies look for in quant jobs these days tend to be ML skills these days, as opposed to traditional TS skills. PhD in CS seems to be in most demand as opposed to PhD in math or other stuff. $\endgroup$
    – confused
    Commented Aug 8, 2020 at 17:19
  • $\begingroup$ At the end of day, the best model to use depends on the problem at hand. If the true relationship is linear with guassian conditional distribution, the best model may still just be a linear regression not a fancy ML model. Don't quote me on this, but I would imagine if you want to predict steps ahead, a traditional model may still be the best since you have a mathematical formula that allows you to predict ahead. For ML, you generally need to put in the observed data as you move ahead in time - unless you structure your response as a x day ahead sequence. $\endgroup$
    – confused
    Commented Aug 8, 2020 at 17:23
  • $\begingroup$ thanks, the problem at hand is volatility forecasting. financial asset return volatility would be good $\endgroup$
    – develarist
    Commented Aug 9, 2020 at 15:47

This largely depends on your setting and the available features.

You can include further information into classification or regression algorithms by providing the model with additional features such as the daily, weekly, monthly returns of previous periods and eventually also use these to create more features such as measures of volatility, mean-reversion or other aspects you would include in a "traditional" approach.

Benefits of machine learning algorithms (especially neural networks such as LSTMs or other RNNs) are, that they tend to be really fast and still offer a comparably good performance to many sophisticated option pricing models.

  • $\begingroup$ ok i could incorporate features besides autoregressively using preceding volatility estimates, but how well has this worked in publications? Before even going to other features, how have distinct machine learners compared against one another when they only use preceding volatility estimates (autoregression) for features? Which ones perform better for forecast horizons longer than 1 day ahead? $\endgroup$
    – develarist
    Commented Aug 1, 2020 at 2:40

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.