Recently I started reading "volatility trading" by Euan Sinclair. On the chapter about forecasting RV he talks about the GARCH model, but he kind of implies that its more stuff for academics, not so reliable in practice. The main criticism is that in order to define the parameters for the equation you need to do a lot of curve fitting. My question is, anyone here is using models from the ARCH family to a certain degree of success? Or should I look elsewhere without wasting too much time on things like ARCH GARCH EGARCH TGARCH OLIGARCH etc.. (ok the last one I made it up) ?
I suspect, without any proof, that a lot of the academic stuff is bullshit and you shouldn't follow it too closely, but just look at it for ideas.
Any auto-regressive process is just doing on a monotonic basis what our eye does instinctively: it interpolates from what has happened to extrapolate onto future points. We do the same thing when we hear a headline and then look at price/volume action, and think, "Gee! That's sure an over-reaction!" [or whatever-the-hell we go on to think...] The periodicity we need to follow (it is important to remember) is a massed result of a million individual, *mostly* uncoordinated trading agendas, and so is not a static thing. (If it were, we could just use the OLIGARCH model, and "Whammo!" out come the perfect results! ) My thoughts? Having the AR of auto-regressive in your model is a must. But from there? The mechanics (despite all the self-agrandising labels and artificial differentiation) are basically the same. And from there, the magic question, "But how long!????" for the look-back. My answer to that? THAT is where the fun is. Be testing; be looking; all the time.
Sinclair's book is awesome and he's working on a second one more relevant to retail level traders. Consider contacting him directly, he's active on twitter and quora and i think has his email posted somewhere on the two
@tommcginnis yep in the back of my mind I also suspected it was just another "glorified moving average", however there is one thing I want to believe (otherwise there is really no point in playing this game), which is that predicting the behavior of a million people should be somewhat easier than predicting the behavior of a single individual. Put together all our cognitive biases, the influence of the current media narratives, fight or flight reactions.. for sure you cannot predict the next trump tweet, however can you maybe predict the reaction to it? And maybe there is a model out there that has better than 50/50 chance at it? That's why I keep looking for mathematical models rather than discarding everything immediately as a bunch of academic BS. By the way, if you used the OLIGARCH model in the russian stock market you would have made a lot of rubles this year..
Why not learn them? They were made by very smart people who I am sure you can learn a lot from. Each of them try and capture a slightly different feature of how vol moves. You are searching for the "Aha!" moment right? That moment might come from reading one line in the "OLIGARCH" paper. https://vlab.stern.nyu.edu/en/ this website calculates them for you. GARCH is better than historical vol and worse than implied vol for liquid assets.
If you price vol off of SPX 1-5 month ATMF term structure, and then compare/price via GARCH(1,1) fit with last two years much more heavily weighted, you'll see a close to constant premium (bias) across those tenors. This suggests that at least a few people are using GARCH in practice.
GARCH is literally a type of moving average. It's most basic form is the EWMA which was pioneered for RiskMetrics. It's less about being right more than 50% of the time and more about learning to cut your losers early.