AI and Arbitrage

As an architect of financial software, I am always interested in arbitrage opportunities. Arbitrage opportunities are a way to profit nearly risk free through market inefficiencies across assets or across exchanges.

For a simple example, apples at the farmer market are selling for $2 while at the grocery store they are available for $1.80. One can profit by purchasing the grocery store apples for $1.80 then selling them for $2 at the farmers market.

In actuality, things quickly become more complex. For the arbitrage of apples to work, one would need to ensure the farmers market had enough demand to support the extra apples for sale. Another concern is for one or both markets to learn of the price difference — the farmers market apples may go down in price or the grocery store apples go up in price. The opportunity for these type of scenarios are often very short.

In the recent past, high speed trading systems would look for arbitrage opportunities across assets and markets — taking advantage of the small window where a nearly risk free profit could occur. While that is often profitable, the bigger opportunity is knowing where the arbitrage opportunity actually exists.

For example, buying USD (US Dollars) with EUR (Euros) to buy JPY (Japanese yen) with USD instead of buying JPY with EUR directly so as to obtain more JPY with the same amount of EUR. When this type of opportunity occurs, it could be the result of a problem with the EUR, the USD, and/or the JPY. However, it takes time to do the research.

The window of opportunity will close quickly. The reason high speed trading exists is to take advantage of trading opportunities that are impossible for human traders. These type of trades often occur simultaneously.

Furthermore, an AI system may discover a better opportunity that the actual arbitrage. Using our example of USDEUR to JPYUSD we may find that the JPY is actually stable in price against the USD but the EUR is the one taking off. Thus instead of making the arbitrage trade, we may decide to investigate the best ways to sell EUR.

In some cases humans can predict where the discrepancy is occurring due to experience and research. However, according to the Toronto Destruction Lab, an AI Research Group, “prediction is done better, faster and cheaper by machines.”

My team actually built a similar system to the one discussed above. Using Machine Learning, we discovered that an arbitrage opportunity is almost one sided. Emphasis on almost. Perhaps more enlightening is the fact that all of the assets we found present in an arbitrage opportunity have some degree of inefficiency.

Going against the Efficient Market Hypothesis, it appears that prices are almost random, therefore markets are almost efficient. Sure, prices move based on events. However, that is far from being able to predict their direction or height.

Benoît Mandelbrot, the Mathematician famous for Fractals, in his study of markets concluded, “People want to see patterns in the world… So important is this skill that we apply it everywhere, warranted or not.” His conclusion was that market prices from the 1800's Cotton Trade to present day are actually more random than anyone ever suspected.

Believing that an AI based system could predict any market is science fiction. Trying to predict market prices is just a little less difficult than predicting if a coin will be heads or tails. AI can measure inefficiency better than humans in spotting arbitrage patterns, but knowing when to enter and exit such opportunities is still hard, if not impossible to determine.

The problem is that predicting near random events is probably impossible. James Surowiecki in his 2004 book, The Wisdom of Crowds, introduced his findings that illustrate how large groups have made superior decisions in psychology, behavioral economics, and others fields. However, this theory has failed in making predictions.

Barry Ritholtz, wealth manager and columnist, explained that prediction markets failed spectacularly in trying to guess the outcomes of events such as the Greek referendum, the Michael Jackson trial, and the 2004 Iowa primary.

Humans can spot arbitrage patterns, are pretty good at discovering long term investments (ex: Warren Buffet), but are lousy at making predictions. It appears AI will be no better at this.

Combining AI with High Speed Trading has provide never before realized opportunity in arbitrage. Over the next few years, these AI based systems will improve. However, it is unlikely if an AI or any other system will be able to predict even slightly better than random.