Forums  > Pricing & Modelling  > forecasting volatility with high frequency data  
     
Page 1 of 1
Display using:  

rickyvic


Total Posts: 246
Joined: Jul 2013
 
Posted: 2021-04-14 07:22
Hey guys

I was hoping anybody could shed some light on any decent framework to forecast volatility at seconds frequencies and/or potentially with irregularly spaced time series.

Some sort of garch with independent variables too.
The literature is massive and messy any pointers appreciated


"amicus Plato sed magis amica Veritas"

ronin


Total Posts: 684
Joined: May 2006
 
Posted: 2021-04-14 14:41
IMHO, garch and market microstructure don't live in the same universe.

I'd look at market microstructure. Forget garch on those timescales.

"There is a SIX am?" -- Arthur

chiral3
Founding Member

Total Posts: 5221
Joined: Mar 2004
 
Posted: 2021-04-14 16:02
microstructure and sentiment. garch was good 10+ years ago.

Nonius is Satoshi Nakamoto. 物の哀れ

EspressoLover


Total Posts: 486
Joined: Jan 2015
 
Posted: 2021-04-14 20:05
To a first order approximation, high frequency volatility scales within the arrival intensity of sqrt(trade size). You can certainly get fancier, but 4 times out of 5 that approach will work good enough.

Good questions outrank easy answers. -Paul Samuelson

rickyvic


Total Posts: 246
Joined: Jul 2013
 
Posted: 2021-04-15 09:32
Sure thanks.

I was looking for a modelling framework.
I have many orderbook indicators I can use (including arrival intensity thanks @EL).

Good thing about garch/SV type (garch works for daily/weekly not intraday I know) of models is that they ensure the time varying vol is positive and they preserve the meanreverting nature of volatiliity.

I could build a regression dVt = f(Xt-1,...,Xt-p) where Vt is the conditional vol, but I dont think it is a good solution.
Some use squared returns as a proxy (not sure about that either).

Literature is centered on compressing intraday data into daily, which is useless.







"amicus Plato sed magis amica Veritas"

ronin


Total Posts: 684
Joined: May 2006
 
Posted: 2021-04-15 10:27
> (including arrival intensity thanks @EL)

That is a modelling framework.

Price processes simply aren't continuous Brownian motions on those timescales. Pretending that they are won't help.

Any meaningful modelling framework has to include the interaction between orderbook dynamics and market orders. The var ~ qty thingy is the limit when the orderbook change rate is significantly slower than the market order arrival rate. Which may or may not be relevant, depending on what you are modelling.

In terms of the literature, the key words to search for are "market impact" instead of "volatility". Look at CFM's library of old papers to start with. And implement a decent vwap.

"There is a SIX am?" -- Arthur

chiral3
Founding Member

Total Posts: 5221
Joined: Mar 2004
 
Posted: 2021-04-15 11:58
As someone that has used GARCH and all it’s various incarnations as part of systems in daily reading for the past 15 or so years it has been my observations that it is not terribly useful when things matter, particularly in the last 5 or so years.

Nonius is Satoshi Nakamoto. 物の哀れ

rickyvic


Total Posts: 246
Joined: Jul 2013
 
Posted: 2021-04-16 08:58
Ok cool thanks @chiral3, I used garch and sv extensively to target risks with a decent success on slower strats, trend following, fundamental, etc...

@ronin Appreciated your pointers, I am well aware of the literature on market impact and microstructure and I understand where you are coming from, although my specific case is a little more complex because of the otc/fragmented nature of it.

I will probably go for an easy approach pooling different ordebook orderflow indicators within a linear framework then maybe go back and refine.

"amicus Plato sed magis amica Veritas"

chiral3
Founding Member

Total Posts: 5221
Joined: Mar 2004
 
Posted: 2021-04-16 13:06
And I don't know, maybe it's because I live in the US, and I trade a good amount of US (maybe 70% versus 30% ex US), and maybe because of Trump, and US policy (fiscal and monetary), and the business environment, and COVID, etc etc. but the value in other sources of data, derived from non-quant, has become much more powerful in recent years. Doesn't apply to your time scale but the usual suspects out to t+30 are good indicators of what has happened already. It's like they are either not predictive or so reactive that the tx dominates.

Nonius is Satoshi Nakamoto. 物の哀れ

EspressoLover


Total Posts: 486
Joined: Jan 2015
 
Posted: 2021-04-16 20:22
> but the value in other sources of data, derived from non-quant,

Interesting. Any chance you could share more color on what kind of non-quant data you're looking at or how you're integrating it with traditional market data? Certainly understandable though if mum's the word because it's secret sauce.

Good questions outrank easy answers. -Paul Samuelson

rickyvic


Total Posts: 246
Joined: Jul 2013
 
Posted: 2021-04-19 07:21
I was intrigued too. I would think of news and ECO calendar.
We all very concerned trying to forecast the first moment, but second moment gives a huge edge too with position sizing.

"amicus Plato sed magis amica Veritas"

ronin


Total Posts: 684
Joined: May 2006
 
Posted: 2021-04-19 08:52
Reddit is apparently all the rage these days.

I get a new person trying to push a Reddit monitoring tool on me pretty much daily.

"There is a SIX am?" -- Arthur

gagyang


Total Posts: 5
Joined: Jun 2020
 
Posted: 2021-04-23 09:20
Could you elaborate a bit more on what you mean by a decent vwap? The few parameters I can think of in vwap would be amount of volume to look back? Possibly adding heavier weight to recent samples and adding time limits to the sampling as well? Normalize for market depth?

Also, from a high level overview, what would be a evaluation method for a “good” vwap? To see if it’s a better predictor of future mid price than simply using current mid price?

ronin


Total Posts: 684
Joined: May 2006
 
Posted: 2021-04-23 16:26
Well, this might sound facetious, but here it is:

A decent vwap does a decent job of, you know, capturing vwap.

A vwap isn't a predictor of mid price, it's a predictor of vwap. And, if it does that, there is an added bonus - it then also does a decent job of predicting intraday quantity and intraday volatility.

These things that chase yesterday's vwap are probably not a great choice. You want a stable estimator, you don't want to keep chasing your own tail all the time.

"There is a SIX am?" -- Arthur

chiral3
Founding Member

Total Posts: 5221
Joined: Mar 2004
 
Posted: 2021-04-23 16:39
I'll get a little polemical. Years ago I was looking to throttle a rebalancing strategy as a function of vol and found that nothing that worked the previous decade was very helpful. The better solution involved forecasting vol and a simple sentiment indicator was better both in and OoS.

Caveman view via economics: short comings in both the older rational / expected utility school and the newer behavioral school. The mixture of the two is still flawed, but better at explaining certain observed items (e.g., ERP). So my read on this is a paraphrase of things I've said differently for years, esp since 2008. People weight certainty highest and the things that are certain today are driven mainly by sentiment, news cycle, virus, global events, surprise, fiscal/monetary, etc. OK, timescales matter. People define HF differently. I admittedly trade at time scales from minutes to years. With that caveat, and I completely agree with comments re vwap and sqrt(size), markets have been overweighting nearer-term events with higher implied tails than in the past. It's silly to talk about the disconnect with fundamentals, but that was one of many canaries in the mineshaft. So my secret sauce isn't so much the signals themselves, which most of you would find rudimentary, but all the really practical smoothing and filtering and winsorizing that happens. 10 years ago I had my egarch models and my other inputs, like dealer gamma, volume, things derived from the options markets, ..., but those things just don't work for forecasting vol anymore. Again, caveat emptor on timescale.

Nonius is Satoshi Nakamoto. 物の哀れ

JTDerp


Total Posts: 83
Joined: Nov 2013
 
Posted: 2021-04-23 17:12
Not trying to hijack, but chiral: you mention dealer gamma being ineffective for forecasting vol...has any research by Nomura or the squeezemetrics.com site been of any value to you in the past few years?

"How dreadful...to be caught up in a game and have no idea of the rules." - C.S.

chiral3
Founding Member

Total Posts: 5221
Joined: Mar 2004
 
Posted: 2021-04-23 18:04
I have a bunch of sources for data. Haven't use squeezemetrics so can't comment and have a relationship with Nomura but not in this regard.

Nonius is Satoshi Nakamoto. 物の哀れ

gagyang


Total Posts: 5
Joined: Jun 2020
 
Posted: 2021-04-24 03:25
> A decent vwap does a decent job of, you know, capturing vwap.

> A vwap isn't a predictor of mid price, it's a predictor of vwap. And, if it does that, there is an added bonus - it then also does a decent job of predicting intraday quantity and intraday volatility.


Not sure if I am following what you're saying correctly, but if you mean using vwap as the model target, why not directly measure the weighted variance or variance?

ahgt_123


Total Posts: 25
Joined: May 2020
 
Posted: 2021-04-24 22:53
@chiral3 by sentiment, dont say that you meant twitter sentiment?

Once I downloaded 1yr of twitter data for something from an online tool, took about 6 hrs.
Out of curiosity looked at a couple of days of tweets (~300/day) and came to the conclusion
that there is no way that crap can predict anything let alone markets.

Please dont tell me that there was gold in that shit.


caseywh


Total Posts: 1
Joined: Apr 2021
 
Posted: 2021-04-26 00:40
Once you have this setup you could probably compare it to a model of UVXY aum.

UVXY King.

ronin


Total Posts: 684
Joined: May 2006
 
Posted: 2021-04-27 10:28
> why not directly measure the weighted variance or variance?

It's kind-of the same thing.

Generally, traded quantity is supposed to be more 'natural', in the sense that it is the aggregate of a mindless crowd. A bit like molecules in a gas. You have no idea what any one of them will do individually, but the gas is completely predictable and simple.

The orderbook is supposed to be populated by somewhat intelligent human beings. So it's less 'basic', in some sense.

But you are free to model what ever you want. If you get an edge modelling the variance directly, do you really care if somebody thinks you are doing it wrong?

"There is a SIX am?" -- Arthur

bandi_np


Total Posts: 11
Joined: Nov 2020
 
Posted: 2021-04-28 07:50
@ahgt_123 I remember two papers (https://arxiv.org/abs/1307.4643 and https://arxiv.org/abs/1403.1715) that discussed the predicitve power of Google trends and how the set of keywords used for analysis can have an impact. It was 7 years ago so it might change now.

doomanx


Total Posts: 115
Joined: Jul 2018
 
Posted: 2021-04-28 09:18
Late to the party but I think the papers by Andersen and Bollerslev on this topic cover a lot of the right ideas (or at least are a good starting point, although they're not trying to forecast VWAP).

did you use VWAP or triple-reinforced GAN execution?

rickyvic


Total Posts: 246
Joined: Jul 2013
 
Posted: 2021-04-29 07:31
To be honest it is a lot simpler guys.
When you need a second moment in your trading logic it means you need to know by how much in a certain horizon price is going to be expected to move without knowing in which direction.

That way you can correctly size your bet, whatever you are doing

Ok so if you take the squared returns it is a fine approximation of variance, then you find ways to predict this variable by using whatever you find useful given the horizon.

The problem with garch/SV models is that they work on long horizons because they capture the meanreversion and other characteristics of price and conditional volatility distribution.
Realized vol has a lot of insights in it, however the very nature of the estimators implies asymptotics, which works on a lot of data only, so aggregating realized/instantaneus vol estimates on horizons shorter than hourly makes no sense.
Some say 5 minutes but I had no luck isolating jumps and estimating vol using that.

"amicus Plato sed magis amica Veritas"
Previous Thread :: Next Thread 
Page 1 of 1