There is an old expression in computer programming: “Garbage in, Garbage out.” The point of the expression is that if you program logic based on bad data or data derived from the wrong context, that the results will not be what you are looking for. In the world of trading cryptocurrencies, the data supplied to machine learning algorithms that determine where to price individual assets might be very unreliable at times, and the data for trading those assets might be even worse (as seen above when BitCoin Cash moved up 2.5 times in value and then back down in less than 3 days).
With that said, there is a lot of buzz around the crypto universe about quantitative hedge funds entering the space and about the use of advanced techniques. One example is that Bitcoin News published a story about the use of Artificial Intelligence in crypto investing. It features an interview with Guy Zyskind, the CEO of Enigma, which offers a data marketplace which hopes to become the foundation for such endeavors.
While I agree wholeheartedly with Guy’s thesis that quantitative trading firms have helped improve liquidity and compress spreads in the equity markets, and also that the firms he mentioned (one of which I spent over 5 years at) have intelligently utilized machine learning and quantitative techniques. The problem with this article, however, is that it is wrong to conflate the use of such techniques in asset selection, portfolio construction and trading. Those are three separate disciplines which require very different quantitative tools and data, and Enigma, like most FinTech firms and Funds in Equities, is focused on asset selection using their alpha generation tools. To be clear, finding alpha is extremely important to funds and investors, but my point is that the tools are different and so is the data.
The article continues the confusion by alternatively discussing trading “bots” and hedge funds in a mashup. In reality, most of the automated market making firms which provide the liquidity that helped collapse spreads are brokers, not funds. Those firms do use machine learning and other quantitative tools for trading and risk management, but those are quite distinct from the type of data and techniques used for alpha generation. To understand these differences, readers can review a previous article I wrote, that went over the different elements of fund management that could benefit from quantitative techniques, “What is Quantamental Anyway?”.
There are many platforms in the equities world focused on alpha, including 100s of alternative data providers, data integrators, artificial intelligence firms with methods of interpreting that data, and many vendors providing access to that data. There is even one firm, Quantopian, who built a platform that anyone can use to find alpha, and uses those “crowdsourced” algorithms to trade. The problem, however, is that capturing alpha is not that simple. Most of the funds that have consistently achieved superior performance have also invested in integrated quantitative processes to build and trade their portfolios. I don’t want to “pick on” Quantopian, but, their advice to their own community includes the following statement: “By starting from a universe that takes volume, slippage, and liquidity into account, we can avoid wasting time during the development of trading signals.” This is very troubling, as it seems to group a large group of stocks together for trading costs and seems to oversimplify the large variations in trading costs that can derive from different market conditions. Perhaps this might explain why they have had some performance difficulties this year…
That said, in equities, as compared to cryptocurrencies, trading is relatively straightforward, with consistent bid/offer spreads, connected markets, and many tools available for traders. In the Cryptocurrency world, however, it is a different situation. That is why we started CoinRoutes, as a platform to help traders understand the market and route orders optimally.
We realized that this market is both fragmented and uncoordinated, and very hard for traders to get a handle on. Markets are usually crossed (meaning that the high bid is above the lowest offer), and order books are hard to read for traders due to the ability to post quotes for a tiny fraction of a full unit.
To put this in perspective, consider a simple trading situation. Let’s say that a fund wants to buy 1000 Ethereum priced in BitCoin (meaning buy ETH and sell BTC). The first thing a trader might do is look at the bid offer spread over the past day. CoinRoutes can show the trader the consolidated bid and offer across 7 different exchanges, filtered to only look at orders over a certain size. For a day earlier this week, filtered for orders greater than 20 ETH, this is what that looks like:
Notice how the system also shows the percentage the market is crossed and the relative market share of best bid and best offer over the day. This alone, makes it clear that a fund trading this pair would need access to most, if not all these exchanges. To explain why the filtering function is so important, however, we should look at the unfiltered view. Notice how in the figure below, when all the small orders are included that the market is almost always crossed, with the majority of the time being large (defined currently as by more than ¼ of 1%):
In this example, if the fund has determined, based on AI or other quantitative techniques, that it should buy Ethereum and sell BitCoin, then they will need to trade in this fragmented market. At that point, the use of a Smart Order Router (SOR) would likely help them capture more alpha in the trade. As an example, I utilized our software to simulate an aggressive sweep of the market to buy 1000 Ethereum vs BitCoin. While this is a large order (over $300k), at the time I wrote this, an aggressive liquidity taking strategy using the SOR would have saved 0.54% even compared to one of the largest exchanges:
Next, I simulated 1/10th of the order, which at 100 ETH is just over $30k, but even then, the SOR outperformed a single exchange at that time by 0.5%. This is likely due to the shifting between exchanges and may represent a higher number than average, but it was based on live market data:
The point of this example is to show the importance of trading, but it is equally true that the data itself must be understood. Crossed markets, in particular, are indicative of potential structural issues, and could confuse machine learning systems unless well modeled.
In conclusion, we believe that due to the fragmented and uncoordinated nature of the crypto asset markets, it is likely that funds who understand, and have tools for managing market microstructure, will outperform those that do not. The nature of the crypto markets means that such tools are likely much more important trading BitCoin, Ethereum, BitCoin Cash, LiteCoin, or any of the newer “Alt Coins” than when trading equities. Thus, while Enigma, and the Artificial Intelligence strategies that it facilitates, may well represent a major advance, it is not a panacea. Even if investors were able to isolate which coins to buy and sell at any given moment, unless they are capable of doing so efficiently, they will not perform up to their expectations and won’t capture the alpha they expect.