About Order Book Models

B

bbt80

Recently I've started to research about order book models and HFT (incidentally).

Note: Although I don't intend to trade this way, I need to implement proofs of concept to test hypotheses (as I won't backtest tick data). So I'll trade occasionally. I know I contradicted myself.

My approach is a balance between explanatory and viable models.

I've already read books and papers about market microstructure and other technical details. Right now I'm improving my theoretical background (from measure theory to numerical methods) and programming skills (Julia and C++). I've gathered some papers and books about order book models and I'm planning to store tick by tick data next year.

I'll keep it simple: Can I infer this parameter in 50 ms (just an example)? If no, then it's useless. Deep learning? I don't think so. Just the old and plain stochastic processes, Laplace/Fourier transforms, analytical stuff in general.
 
Last edited:
My background is in Mathematics, Statistics, Data Science and I have basic/intermediate programming skills. I've been trading/investing by myself since 2003 (not for a living).

My goals here are to:
  • share my advances in this area of research
  • share my thoughts about related stuff
  • pick up tips from experts
  • get opinions
  • socialize
Does it pique some interest from you? Or at least some curiosity?
 
nope. even if they were slow. they would still win due to the exchange arxhitecture as well as low fees and rebates.

they build order books with ticker plants and manipulate cash prices in jersey to change or set up their trades within a millisecond in chicago.

imagine you create liquidity in cash n futures getting paid on both because you can move the mkt and have priority.

if you monitor and create price levels and need an extra 10 milliseconds to cancel orders on the next level before the market moves all you do is keep selling into the buying as you cancel above sell orders l. this is them taking a loss.

you will never have the capital or connections literally and figuratively to beat them. you are starting 20 years too late.
 
hawkes process based microstructure models are the most realistic . I do not think you can win at HFT tho. entrenched players are too fast and deep pocketed . think about how the microstructure scales.. from hawkes to fractionally integrated CIR diffusion
 
stochastix said:
I do not think you can win at HFT tho. entrenched players are too fast and deep pocketed . think about how the microstructure scales..
More...
And not to mention two important issues:
1) Fees and Commish: you can create a winning trading model/system that is profitable BEFORE fees and commish...and shows a great equity curve upwards. Then, when you apply fees and commish, the equity curve is ....STRAIGHT DOWN. A loser.
2) Using Level 2 data ? You better create some filters to reduce the impact of spoofing. Otherwise, your models will get blind-sided hugely by these "head fakes".
 
It's not hopeless as some posters stated, but it's not easy either.

The general idea is that you can't beat high frequency players because of technology costs, but you can participate in the game when they are not interested. So what you want is to provide liquidity at the holding horizons or imbalances that are beyond that of the traditional HFT players. Remember that these guys do not like taking on any meaningful risk, so any 2+ sigma events in liquidity will not be readily adsorbed.

There is some good and some bad to this business model. If you take meaningful risk (unlike true HFTs), your Sharpe ratios would still be much higher than the market though you're still going to have losing days/months. Also, pnl per trade value wouldn't be as tight as that for market makers, so you are less impacted by trading costs. The bad is that this approach usually requires a rare combination of the alpha skills (market knowledge, statistical methods etc), with technology acumen (ability to build systems that take advantage of these opportunities) all while operating in a rather capacity constrained way.
 
syswizard said:
And not to mention two important issues:
1) Fees and Commish: you can create a winning trading model/system that is profitable BEFORE fees and commish...and shows a great equity curve upwards. Then, when you apply fees and commish, the equity curve is ....STRAIGHT DOWN. A loser.
2) Using Level 2 data ? You better create some filters to reduce the impact of spoofing. Otherwise, your models will get blind-sided hugely by these "head fakes".
More...
I implemented a model that optimized fees, commisions, taxes, spread transitions, diurnal variation, etc. Even with all that, 60 microseconds response time when cancelling and reinserting orders is too slow, your lunch will be eaten if you are trying to peg orders to the BBO. The best exchange to do this on is IEX since they have a protected order, and now that the SEC approved the D-limit that might work but when I did this last summer the order type was not available.
 
You guys are great! That's what I expected to hear from you as I didn't have the time to explain more about my “strategy”.

In the next days I'll clarify things, but, for now:
  • I am a student and this is just a theoretical work (but, as I said, I'll test it in the wild)
  • I'll not trade in the US (if I can compete everywhere in the world, why would I choose the hardest championship?)
  • I will not try to occupy the space of HFT and market makers, but I'm interested in modeling their behavior
  • From my perspective, the order book is like a cell in a tissue (down-top approach)
Mental exercise: Imagine big international banks and funds investing in some market (not the US) index and you can see their (the brokers) activity. Almost every single order. There're a couple of ETFs based on that index and some other related. The index is concentrated in a dozen stocks. There's futures and options markets based on that index. The big guys need to hedge their exposition to the exotic currency too. Everything is related. It's like an orchestra trying to keep up with the pace of a crazy conductor. They make a lot of small mistakes (almost imperceptible). But there're times they make big ones.

Some people coined the expression "guerrilla trading" sometime ago. I think it is aggressive. I'm not even a trader (let alone an opportunistic one). I'm just a guy with a bad taste for music. But let me just start with a simple order book model.
 
thats a good way to put it . i once saw a video describing it as some guy trying to keep like 60 plates spinning at once by continually adjusting each one.. occasionally they wobble and all smash at once
 
bbt80 said:
You guys are great! That's what I expected to hear from you as I didn't have the time to explain more about my “strategy”.

In the next days I'll clarify things, but, for now:
  • I am a student and this is just a theoretical work (but, as I said, I'll test it in the wild)
  • I'll not trade in the US (if I can compete everywhere in the world, why would I choose the hardest championship?)
  • I will not try to occupy the space of HFT and market makers, but I'm interested in modeling their behavior
  • From my perspective, the order book is like a cell in a tissue (down-top approach)
Mental exercise: Imagine big international banks and funds investing in some market (not the US) index and you can see their (the brokers) activity. Almost every single order. There're a couple of ETFs based on that index and some other related. The index is concentrated in a dozen stocks. There's futures and options markets based on that index. The big guys need to hedge their exposition to the exotic currency too. Everything is related. It's like an orchestra trying to keep up with the pace of a crazy conductor. They make a lot of small mistakes (almost imperceptible). But there're times they make big ones.

Some people coined the expression "guerrilla trading" sometime ago. I think it is aggressive. I'm not even a trader (let alone an opportunistic one). I'm just a guy with a bad taste for music. But let me just start with a simple order book model.
More...
I like where you are going with this. This was my approach when I tried implementing the Guilbaud-Pham 2013 model. I used the zero-intelligence limit order book matching engine simulator implemented in a project called parity on github. When I ran my implementation against the simulated market, I got results which matched the statistics of the phd thesis results which I based my work on pretty much perfectly. I'm talking the shape of the PnL distribution, the mean profit per share, etc. So after this step, when I took my implementation to the real market I found out that actual agents operating in the market are operating logic that is timed perfectly across many exchanges.. they can cancel an order on one exchange.. and then knowing they did that, send a perfectly timed cancel and replace order to another exchange because they know anyone operating a pegging strategy will necessarily act in a certain way due to latency and the deterministic response of the other participants. A theoretical physics justification for this type of thing being commonplace can be found at http://matpitka.blogspot.com/2020/12/new-evidence-for-zero-energy-ontology.html this guy was a PhD student of Wheeler (student of Einstein) who got kicked out of the university of Helsinki because he dared challenge the orthodox interpretation of the mass=energy relation.

imho, this is the best way to understand microscopic hft -> macroscopic.

From Hawkes-type processes to stochastic volatility
"We introduce a Hawkes-like process and study its scaling limit as the system becomes increasingly endogenous. We derive functional limit theorems for intensity and fluctuations. Then, we introduce a high-frequency model for a price of a liquid traded financial instrument in which the nearly unstable regime leads to a Heston-type process where the negative correlation between the noise driving the process of the instrument and the volatility can be viewed as a result of high variance of the sell-side order arrivals."





there was a project i Read about where they were using generative processes to to generate Hawkes-like processes which modeled the reactions of actual market participants. This would be a viable way to simulate a true HFT like environment I think, but I did not go that route.

I actually have a lot of Java code to implement this stuff that I am not using anymore. If you are interested, ping me
 
You must log in or register to reply here.
Back
Top