Thread by @TerribleQuant

For anyone making an HFT strat, you need a simulator to see when you get filled. You can use Kalman queues, multi-queues, sim matching engines and they are all cool but usually don't properly capture effects like adversity. Then there are stochastic approaches...
1/n

This would be like simulating the poisson process. This sucks even worse as it utterly and entirely ignores adversity. At least sim matchers give a try, although not a great one and an easy NN will trounce it. Plus stochastic approaches don't use historical data...
2/n

Other methods involve procedural OB sims which basically just take the data and add a bit of stochastic overlay to the historical OB, this assumes (most of the time) that you get filled when midprice crosses your bid/ask which is just wrong because it neglects...
3/n

bounce (bids and asks getting executed without crossing the midprice). The best approach is the simple ML approach. This is where we use a machine learning model to predict whether you will get filled based on the data without assumptions at all. There are ways to improve...

4/n
This with stochastic methods to regulate hetroskedasticity and kurtosis in your fill SD that usually won't get picked up by an ML algorithm, but we will ignore that for now since it's a bit of work for not much extra performance. Anyways back on topic...
5/n

We first start by labeling our data. What are we even predicting? One simple approach is to just label the time until an order gets filled, and from there you can infer the conditional volatility for stochastic sims after the fact. For this, you need some LOB data and some...
6/n

trade data. This can be acquired if you get approved for it from Binance for no costs (may want to be VIP or they will ignore your email), or just ask me and I'll load some in a drive for you. Now that we know what we are predicting what are our features? The first...
7/n

and largest driver of variance will simply be offset, and as many familiar with basic HFT theory volatility dependent offset plus a bit to do with capitalization and tick sizes usually determines bid-ask spread. The volatility-dependent offset model is most appropriate...
8/n

when dealing with cryptocurrencies, forex and equities have different microstructural considerations (forex is OTC, and equities have VERY different behaviors between large and small caps). Small caps are more multi-fractal, but large caps are strangely jump stochastic...
9/n

Back on topic. We engineer offset from midprice (obviously), but also microprice. This is just a simple prediction of the midprice 1 tick from now, you can do this again for 2 ticks as well. Weighted midprice is good as well, but for reasons I won't discuss inadequate...
10/n

to be used as a primary feature. Basically, you get the convexity you should really when factoring liquidity with micro price, and you get a diagonal with weighted (liquidity leaving doesn't mean fair value always changes), and obv flat with normal midprice...
11/n

Next is skewness of the orderbook, this is standard and is usual for books. You can also use the time of day. There are about 5-6 cycles of depth in a sin wave like form for illiquid cryptos, slightly smoother for liquids. This is stationary so you can use FFT since you...
12/n

only need frequency, not time unlike DWT would give you. Now we have the most important factor to decide your model's success live -- Adversity. This is just the fact that you will get filled way more when you have no alpha. Nobody wants to make markets for an arbitrageur...
13/n

So you need a fast model (unless you have a million-dollar compute budget) and train it on multiple timeframes to predict. Different OB levels will favor different term predictions. Then finally include microstructural models like VPIN and lambdas that will also pick...
14/n

up on adversity to some degree, although not very well. For those who are beginners, rolling STD is not the only volatility and OB based volatility metrics are great and just other volatility metrics that aren't even OB based. There are also models to use OHLC or tick...
15/n

data to predict the bid ask spread or the skew. Predicted skew or BAS are great features. That's all on simulating LOs and how they get filled. This is a pretty basic view of the main part of what I use, but hopefully, it is insightful. Share your approaches!
16/16