Forums  > Pricing & Modelling  > mean reversion and ornstein uhlenbeck in trade time  
     
Page 1 of 1
Display using:  

rickyvic


Total Posts: 222
Joined: Jul 2013
 
Posted: 2019-12-19 15:26
Hey guys any pointers at these models for high frequency data?
I am looking at the literature I was thinking of an estimator robust to microstructural noise that works in trade time.
Is this even an important issue or I can do with the normal one?
Found this but I am pretty sure there are better papers to start from
https://arxiv.org/pdf/1811.09312.pdf

"amicus Plato sed magis amica Veritas"

nikol


Total Posts: 1234
Joined: Jun 2005
 
Posted: 2020-01-05 20:58
I found this ref because came myself to the following:
Use of (averaging) filter in trade trigger (MATT) design leads to mean-reverting behavior of Price-MATT. Since everyone is using that you should have observed the collapse of the amplitude or constant change of the (averaging) period. Or the change of other parameters of the filter you use. I would call it secondary (market feed-back) effect.

As it is said "they publish only things which do not work".

Sorry for the late response. It is not HF-style, no (hope not a sin).

rickyvic


Total Posts: 222
Joined: Jul 2013
 
Posted: 2020-01-18 21:41
Thanks for the answer I will look into it... My question is how to capture the jump in mean that happens periodically.
I tried a bunch of things in state space but I can't make it fit, while clearly there is a strong negative autocorrelation of returns.
I am also considering adding a jump term instead of a time varying mean, but it is not picking up the jumps correctly. Maybe still need to check I am doing right.

"amicus Plato sed magis amica Veritas"

rickyvic


Total Posts: 222
Joined: Jul 2013
 
Posted: 2020-01-18 21:50
Microstructure noise does not bother a mean reverting portfolio. Jumps do as it is non stationary in level but locally strongly meanreverting although not necessarily stationary.

"amicus Plato sed magis amica Veritas"

nikol


Total Posts: 1234
Joined: Jun 2005
 
Posted: 2020-01-18 22:19
This one is interesting to look at.
https://www.risk.net/risk-management/1940305/event-risk-modelling-equities

Basically you measure diffusion part and estimate jump contribution at the same time. When Jump comes (beyond quantile X) take it into Jump, but exclude from diffusion.
As third contribution, you can incorporate your overreaction as mean reverting process on top of diffusion within T-period after jump (parameter). But that might be subtle (I guess).

rickyvic


Total Posts: 222
Joined: Jul 2013
 
Posted: 2020-01-21 17:54
I have the same approach also with added scheduled events and so on...
I have fitted a model that attempts to predict the jump component, so first estimate the jump size and probability then try to predict it.
Something is not working on that side as I can predict it but it happens too often. Using B-N jump test and a regression model to predict it.

The jump reversion type is an option too assuming you can measure correctly the jump.

Since it is for a hedged portfolio I am looking for a dynamic beta that adds a degree of uncertainty on whether to hedge or not. That solves the issue of non stationarity a bit as the portfolio is always bounded and mean reverting (with the problem of a time varying beta).

So the way I see is:
1) make a time varying mean capturing the jump
2) detect and model a jump or no jump regime
3) make the portfolio stationary at the cost of re-hedging
4) all of them combined

"amicus Plato sed magis amica Veritas"

nikol


Total Posts: 1234
Joined: Jun 2005
 
Posted: 2020-01-27 17:27
Sorry, lost your response.

Jump = intensity (probability) + shape of amplitude.
Both parameters are function of cut-off (separator) from diffusion and perhaps of some other parameters. Or you can simulate diffusion+jump as a joint distribution as well.

> I can predict it but it happens too often.

Check sensitivity to cut-off parameter - do you cut by calibrated sigma of (Brownian) diffusion or by quantile of (non-Brownian) distribution or you blend (convolute) them..

GARCH or mix of two diffusion processes might give you same fat tails.

Found this :
"Detecting Jumps in High-Frequency Prices Under Stochastic Volatility: A Review and a Data-Driven Approach", by Ping-Chen Tsai and Mark B. Shackleton. (book: "HANDBOOK OF HIGH-FREQUENCY TRADING AND MODELING IN FINANCE").

rickyvic


Total Posts: 222
Joined: Jul 2013
 
Posted: 2020-02-18 13:57
Thank you I will be looking into it.

"amicus Plato sed magis amica Veritas"

Its Grisha


Total Posts: 61
Joined: Nov 2019
 
Posted: 2020-06-18 20:36
I have a related, but perhaps slightly more broad question on microstructural noise. I'm seeing papers like the one in OP going to great lengths to control for microstructure effects like bid-ask bounce.

If I have limit order book data, why bother with trades and trade time at all to define my conception of price and price ticks? Of course there is some alpha information in the trades, but to me it does not make sense to define price based on the last trade. I can use naive order book midpoint, or a formula that takes into account resting sizes. A "tick" is just the changing of this midpoint calculation.

Are the academics using trades for the price process because they don't want to deal with acquiring and processing book data? Or am I missing some other major point here? To me the "midpoint process" seems naturally better behaved than the "trade process".

nikol


Total Posts: 1234
Joined: Jun 2005
 
Posted: 2020-06-18 22:38
Can you by having LOB, projected price at T and trade intensities develop this?
- optimal MM strategy
- optimal order execution path as a combination of Making and Taking
by maximizing E[PnL] vs var[PnL] given utility or price of risk.

Its Grisha


Total Posts: 61
Joined: Nov 2019
 
Posted: 2020-06-19 00:54
@nikol Not without consulting a paper, can you recommend a good one? :)

nikol


Total Posts: 1234
Joined: Jun 2005
 
Posted: 2020-06-19 06:53
1. Most the time it is linked to HJB equation.
2. Predictions - econometrics (physical measure), e.g. "Empirical market microstructure", by J.Hasbroek. Or/and use other instruments (risk neutral measure).
3. MM - follow Avellaneda , Stoikov
4. Optimal execution - follow path of Almgren, add a bit of interaction with the market (via MM style of thinking)
5. Account for Inventory too - follows line of classical Amihud & Mendelson (~1980), but look more advanced techniques.

Combining all of them will give you what you are looking for. To be honest, I didnt complete this full path, just imagined "programma" to follow.

At NP, "University" group/first page there are still few gems worth reading.
F.e.
https://www.nuclearphynance.com/Show%20Post.aspx?PostIDKey=193335

ronin


Total Posts: 610
Joined: May 2006
 
Posted: 2020-06-19 20:38
> If I have limit order book data, why bother with trades and trade time at all to define my conception of price and price ticks?

Strictly speaking, these are different price processes.

The prices you really, really care about are "the price I can get in / out at". That's the order zero, and that's just orderbook.

Trades come in at higher orders, depending on what you are trying to estimate.

Trades alone, without the orderbook, are nonsense, of course. When you move away from liquid stuff, you can easily see things that tick once per day or less. You would be stretching definitions to even call that "a process".

"There is a SIX am?" -- Arthur

nikol


Total Posts: 1234
Joined: Jun 2005
 
Posted: 2020-06-22 13:53
@ronin

> When you move away from liquid stuff, you can easily see things that tick once per day or less.

Why not. If your model is able to predict near zero ticking, then you correctly capture that process.

Its Grisha


Total Posts: 61
Joined: Nov 2019
 
Posted: 2020-06-22 17:52
@nikol Sorry have not checked the thread, thank you very much for your literature recommendations. Will be working through them. Previously have not needed much microstructural thinking but increasingly interested in strategies which consider it.

@ronin So you would agree there is a misalignment with how a lot of academic literature defines a tick? Because in a high frequency context, as you say, the book is what matters.

ronin


Total Posts: 610
Joined: May 2006
 
Posted: 2020-06-24 20:07
> @ronin So you would agree there is a misalignment with how a lot of academic literature defines a tick? Because in a high frequency context, as you say, the book is what matters.

Well, I have read plenty of academic literature about orderbooks.
But, yes, on the whole I agree. Most of it is nonsense.


> Why not. If your model is able to predict near zero ticking, then you correctly capture that process.

Yeah, that is not entirely what I meant. What would that sort of process have to do with any meaningful definition of "price"?

"There is a SIX am?" -- Arthur

nikol


Total Posts: 1234
Joined: Jun 2005
 
Posted: 2020-06-24 20:24
@ronin

> meaningful definition of "price"

Well, let's start discussion ))

EspressoLover


Total Posts: 461
Joined: Jan 2015
 
Posted: 2020-06-25 16:32
@grisha

The continued use of last trade as a metric of price is pretty much just a holdover convention from the bad old days when an electronic LOB didn't exist or wasn't easily visible. A lot of the foundational academic literature on microstructure was done in the 80s and early 90s, when trade prices were the only real historical series available.

Plus I think there's maybe some stickiness from the fact that "last traded price" is a lot easier to explain without getting into gritty details that define mid-price. So, anything civilian-facing, like Yahoo Finance or Bloomberg, will just use the convention that "price" is synonymous with "last traded price".

In practice, I've never heard of a serious HFT operation that treats last traded price as the fundamental price metric. For any reasonably liquid market, mid-price or weighted mid-price seem nearly universal.

Good questions outrank easy answers. -Paul Samuelson

gaj


Total Posts: 115
Joined: Apr 2018
 
Posted: 2020-06-25 17:42
In less liquid markets, last price can be more meaningful than mid price, e.g., far month futures, some less liquid ETFs, or options. The main reason is that those markets typically have few market makers. If they stop quoting for whatever reason, the spread gets super wide, so the mid price process can be wonky.

Some of these products have decent volume. So they're not totally ignored by HFTs. But you won't get very far modelling the price process by itself. You need to look at external signals as well (e.g. use the front month to predict the far month).
Previous Thread :: Next Thread 
Page 1 of 1