High Frequency Trading and Liquidity Crisis

By Dr Arzé Karam, April 2021

The FinTech revolution has brought drastic changes to the way financial markets perform, from currencies, governments and equities to derivatives. Trading now happens in microsecond time frames, and risks are emerging faster and in greater volume than ever before.

In such instances, traditional approaches become too slow to be relevant. In my recent paper with Dimitar Bogoev, entitled “High frequency trading and liquidity crisis”, we focus on the impact of FinTech innovation and the ongoing market disruptions in the wake of the Covid-19 crisis on the major electronic futures markets, that are subject to an increasing frequency of flash crash events (where the price of bonds, stocks, or commodities suddenly plunges but then quickly recovers), e.g. 2010 S&P 500 flash crash, 2015 Euro-CHF peg break, 2016 Sterling Brexit referendum, 2018 VIX Volatility Index.

We propose two metrics to capture the dynamics during liquidity crisis in high frequency markets and outline a novel procedure that allows us to test these metrics across major futures contracts.

During a liquidity crisis, financial markets experience wild price swings due to sudden sell-offs by investors. While electronic markets become increasingly reliant on a high level of liquidity supplied by high-frequency traders during normal times, these traders withdraw from the market at stressful times thus amplifying the initial shock. Consequently, trading environments become more susceptible to these liquidity changes, and this leads to illiquidity spikes to the point of breakdown. When the supply and demand shock induced by the Covid-19 outbreak was added to the picture, we experienced unprecedented events which required the intervention of market regulators to stabilise the markets, as we saw back in March 2020.

We examine whether our metrics identify market disruptions when a market is in a state of a liquidity crisis. We provide evidence on a long sample period that includes days around the 2010 flash crash event and the Covid-19 period from February 2020 to April 2020, in addition to quiet sample periods from 2019. The first metric, namely Quote Volatility (QV), which is based on the rate of oscillation of the best ask and best bid quotations, seems to detect the changes in liquidity supply in a very short period of time. The second ratio, namely Price Momentum (PM), which detects the upwards or downwards of trading price movements, seems to capture the impact of the sell-offs by investors on trading prices.

We use a Bi-directional Long Short-Term Memory (BiLSTM) neural network architecture, which has feedback connections, to estimate our two metrics and a set of liquidity variables. We argue that this approach is flexible and well suited to detect market disruptions based on the data provided.

Our results suggest that the BiLSTM model’s accuracy in detecting the dynamics preceding liquidity crisis using QV ratio ranges between 94 and 95%. This suggests that the model is capturing the patterns when the market is about to crash, versus patterns that are not leading to a crash. Using the PM ratio, the results of BiLSTM model are quite good with an accuracy around 85%. When there is no event, the model correctly assigns a very low probability of a crash happening and it is quite dynamic. It also produces more volatile forecasts highlighting the elevated risks of the market crashing.

Taken together, the results suggest the model seems to capture the increased risks, particularly when the conditions of the markets are extreme, i.e., large order imbalance, abnormal trading volume, and very low level of liquidity. What is very interesting is that the model gives a warning well in advance before we can see the actual spikes of illiquidity that leads to a crash.

Our primary conclusions/contributions are three-fold

Firstly, by identifying the key features that drive financial markets to the point of the breakdown, we show the potential of non-linear machine-learning methods in the field of market microstructure with an accuracy that was not possible with traditional methods.

Secondly, we show the importance of two new liquidity-based metrics in the estimation of machine learning models. The existing theory predicts liquidity crashes are associated with extreme price disruptions, very low level of liquidity in the electronic books, and high trading volume. We argue that the two metrics we propose can detect the combination of these patterns in real time. We assert that market regulators and exchanges should not look at the asset price volatility in isolation; the two metrics we propose are more suitable to identify the risk of illiquidity which predict the low resistance of a financial market to large shocks.

Finally, we note that the metrics and the model we propose are directly valuable for market regulators, exchanges and market participants. The ongoing turbulence across financial markets could raise overall market risk, and shocks can spread across assets, markets and countries with adverse consequences on the real economy. It will also have a significant knock-on-effect on individuals’ pensions, particularly when markets are already fragile and subject to an increasing frequency of liquidity events.

Given the enormous volumes of trade in the electronic futures markets, which will rise massively with the uncertainty around the current economic climate, and the ongoing health crisis, we believe our ratios, once enforced on electronic platforms, are useful in creating early warning signals essential to stopping market crashes as they are forming.

More information about Dr Karam's research interests.