Getting over bid-ask bounce

Question

One property of High-Frequency data is it's subject to bid-ask bounce.

Description : Unlike traditional data based on just closing prices, tick data carry additional supply-and-demand information in the form of bid and ask prices and offering sizes.

As a researcher, it can be an advantage because bid and ask quotes can carry valuable information about impending market moves.

However, bid and ask quotes are separated by a spread. Continuous movement from bid to ask and back introduces a jump process, difficult to deal with through many conventional models.

I am actually working on high-frequency trading project. At the beginning, I wanted to predict the price movement, but because the price is non-stationary and other reasons, it made the problem harder than I was expected. Now, instead of dealing with the price, I deal with the log-returns on the midprice (similar to the standard price) and the features.

Why am I talking of features?

For autoregressive problem, it is well known that RNN-LSTM perform pretty well on a time-series forecasting problem. As the midprices as well as the features are not stationary, then I applied the log-returns on all of them. I think it will be easier to approach the solution of my problem.

However, now, I think I face the problem of bid-ask bounce. How can I get over that setback? Does the log on the midprice is sufficient to face the bid-ask bounce problem?

Answer 1

The joint dynamics of the mid price and the sign of the trades is a complex animal... you should respect its wild nature.

First, I suggest you have a look at this question Optimal Price Metric for High-Frequency Volatility: Executed Price, Mid Price, or Weighted Mid Price?.

Second, you seem to take as well-recognised that "RNN-LSTM perform pretty well [on returns of trades]". Be careful: generally LSTMs (or / and RNN) are used on uniformly sampled time series, and for sure it is not the case of transactions. Serial correlations between the arrival of liquidity (limit orders) and the consumption of liquidity tells us that the sampling if far from uniform: worst that that it contains information... (see the literature of Econophysicists, on Hawkes process in HF Finance, or have a look at L and Sophie Laruelle. Market microstructure in practice. World Scientific, 2nd Edition 2018)

I recommend to read Sirignano, Justin, and Rama Cont. "Universal features of price formation in financial markets: perspectives from deep learning." In Machine Learning and AI in Finance, pp. 5-15. Routledge, 2021. If you read it carefully, it will tell you the the recurrent aspect is not that useful: the state of the orderbook contains already a lot of information about its past .

That was already the message of Cont, Rama, and Adrien De Larrard. "Price dynamics in a Markovian limit order market." SIAM Journal on Financial Mathematics 4, no. 1 (2013): 1-25, and it has been continued in Huang, Weibing, L, and Mathieu Rosenbaum. "Simulating and analyzing order book data: The queue-reactive model" Journal of the American Statistical Association 110, no. 509 (2015): 107-122.

Putting all this knowledge in a end-to-end pipeline should provide what you need.