The Cryptic Use of Quantitative Models in Cryptocurrency Trading
James Eaden smooths his silk Hermes tie with satisfaction. “A little reward for last month’s work,” he thinks to himself. As a luxury car dealer in the mid-1990s in the decidedly Tory neighborhood of Greenwich, Connecticut, Eaden has been doing a brisk business with his new neighbors.
Tucked away in a quiet business park in Greenwich, the new occupants of a nondescript office block have been making salesman such as Eaden and the other purveyors of luxury products in the Greenwich area very rich.
But unbeknownst to Eaden, the people he was selling Porsches, Ferraris and Lamborghinis to inside One East Weaver were, in the mid-1990s, becoming rich beyond the dreams of avarice.
For the better part of the 1990s, Greenwich was the stomping ground of an (ultimately unfortunately named) hedge fund called Long Term Capital Management (LTCM). The fund, which was helmed by two future Nobel laureates had developed (what they believed at the time) to be an ingenious mathematical formula to appropriately price options.
The idea behind an option is simple. For instance, say a stock is worth US$100 today. Now suppose I think that stock is going to be worth US$200 a year from now, wouldn’t it be nice to have the option to by it at today’s price in a year’s time?
If I’m right, I make a tidy profit of US$100 and if I’m wrong, no big deal, it was only an option — the only cost to me is the price of the option itself. The only question is, what should that price be?
Like Baking a Cake with Mathematics
The answer, Robert Merton and Myron Scholes, the co-founders of LTCM, figured, was to be found in a magic mathematical formula. The challenge that Merton and Scholes faced was how to price an option to buy a particular stock on a particular date in the future, while taking into account the unpredictable movement of the price of the stock in the intervening period.
Work that option price out accurately, instead of relying on guesswork, and you’ve essentially found a way to turn copper into gold.
With mathematics short of alchemy, Merton and Scholes reduced the price of an option into this formula,
And if the algebra has you feeling a little queasy, rest assured that in the early 1990s when it first came out, that was precisely what quants like Merton and Scholes wanted.
In order to profit from options, LTCM needed other market participants to be unable to price options accurately — meaning that they would always be overpaying or underpricing options.
In its first two years, LTCM made some serious money by selling options that were never exercised. Because buyers of these options had guessed wrong and LTCM had got it right, LTCM cleaned up.
LTCM also made a killing by snapping up all sorts of securities which their models had indicated to them were mispriced.
By the mid-1990s, Greenwich luxury car dealers like Eaden were doing so well they were able to afford the very cars they were selling.
But because the margins on options lean, LTCM had to borrow money in order to leverage and increase their number of positions. At one point in 1997, LTCM had only US$7 billion in assets, with a further US$126 billion in assets funded by borrowing.
But the staggering amount of debt didn’t keep the quants up at night. Instead, Merton and Scholes rested easily into their thousand thread-count Egyptian cotton sheets, comforted by the belief in their mathematical models which had predicted the odds of the fund’s failure were 1 in 10 to the power of 24 — or virtually zero.
And because LTCM was pursuing multiple (as many as 100 at one point) supposedly uncorrelated trading strategies, the assumption was that even if one of the over 7,000 positions they had went wrong, surely they wouldn’t all go wrong simultaneously?
LTCM’s major business continued to be the sale of options — in particular options that would only be exercised during large fluctuations in the price movements of stocks.
In 1997, the high prices that these options were fetching implied that the markets would be particularly volatile, but LTCM thought this was wrong.
The Trend is Your Friend Until it Ends and Then You Bend
According to LTCM’s calculations, market volatility would actually decline, meaning that the odds of investors exercising their options would be low as well. So LTCM piled the options high and sold them cheap.
And in October 1997, almost as if to prove that Merton and Scholes had ushered in an entirely new era of finance — one rule by quants — they were awarded the Nobel prize in economics.
It seemed as if intellect had triumphed over intuition. Rational models over risk-taking. Non-Gaussian models steamrolling over guesswork.
But even before Merton and Scholes could bask in the glory of their Nobel prize, the Asian financial crisis had already hit in the summer of 1997 — spreading contagion throughout other emerging markets and increasing volatility.
By the summer of 1998, Russia was the next shoe to fall.
In evolution, big extinctions tend to be caused by external shocks — like an asteroid hitting the earth.
The Day After Tomorrow
On Monday, August 17, 1998, LTCM was struck by the financial equivalent of an asteroid.
Weakened by political upheaval, declining oil revenues and a botched privatization, the ailing Russian economy imploded. A desperate Russian government was forced to default on its debts, fueling a wave of volatility that swept through the entire financial system.
Stock markets plunged.
And remember all those low cost options LTCM had sold based on the assumption of decreased volatility? The ones they thought no one would ever exercise? Now suddenly everyone was exercising those options.
According to LTCM’s risk management models, the fund was highly unlikely to lose more than US$35 million on any single day. On Friday, August 21, 1997, it lost US$550 million — 15% of its entire capital.
The traders in Greenwich stared slackjawed and glassy-eyed at their screens. The parking lot at One East Weaver, quickly filled up with chain smokers and those who had just five minutes ago picked up the habit.
By the end of the month, LTCM was down 45% — the only way for the fund to survive was to find a financial white knight. Ultimately, the risk that LTCM posed to the broader financial system forced the Federal Reserve to cobble together a multi-billion-dollar bailout underwritten by 14 Wall Street banks.
The problem with Merton and Scholes’s formula wasn’t that it was mathematically flawed, but that it made assumptions about markets and human behavior and a failure to account for the herd behavior of the market.
Long on Math, Short on History
Another issue was that LTCM’s risk models were working with only five years’ worth of data. Had they gone back just eleven years, their risk models would have captured the stock market crash of 1987. And if they had gone back 80 years, they would have captured the last Russian default after the 1917 revolution.
Which is why the use of highly idealized models in the trading of cryptocurrencies is not only questionable — it’s reckless.
Given that the “true” trading in Bitcoin and altcoins hasn’t been around for more than five years, there just simply isn’t enough data to adopt a purely quantitative approach towards cryptocurrency trading outside of extremely short timeframes.
Take for instance trying to determine volatility in the cryptocurrency markets. To assume that there is uniform volatility across all cryptocurrency exchanges is inaccurate.
Depending on cryptocurrency exchange volumes and order book depth, the deviation in price of even Bitcoin can differ by up to 20% during periods of peak volatility — albeit for very short windows of time.
And there is almost always a difference in the declared price of Bitcoin from one cryptocurrency exchange to another.
Price discovery in the cryptocurrency space is a bit like trying to catch an eel with your bare hands — you can see the eel, you can grab the eel, but within seconds it slips away.
And even standardized models such as the Sharpe ratio (developed by another Nobel laureate, William Sharpe), which is used to calculate risk-adjusted returns are of limited value given the large positive and negative movements in cryptocurrencies.
Cryptocurrency portfolios tend, over longer timeframes, towards kurtosis or “fat tails” — meaning that there is a far higher likelihood of instances where the price of cryptocurrencies fall outside the mean or average.
And because cryptocurrency trading tends to suffer from significant non-linear risks — such as unforeseen and unforeseeable hash wars and hard forks — the use of even other alternative risk-adjusted return calculating methodologies such as the Sortino Ratio and the Treynor Ratio is circumspect at best.
For the most part, these risk management ratios tend to correlate volatility with risk, which is not unreasonable for investments such as bonds, but may be too narrow when applied to alternative assets such as cryptocurrencies.
Correlation Matrices for Cryptocurrencies
Instead, what has been a far more useful measure of volatility has been the correlation between various cryptocurrencies. Between Bitcoin and Ethereum, between Bitcoin and altcoins and between Ethereum and altcoins.
In a bull market, when everything is going up, in particular when Bitcoin is rising, altcoins will generally outperform Bitcoin in percentage terms. And over far shorter time horizons, sometimes condensed to minutes, altcoins have historically had more volatility and higher betas.
In a bear market (similar to the current one), altcoins are still trading at a higher beta, but to the downside. In 2018, Bitcoin fell by about 80%, whereas altcoins fell by 95%. And despite overall cryptocurrency market movements, the correlations between Bitcoin and altcoins and Ethereum and altcoins continues to be very strong, yet altcoins continue to have a higher beta — something completely different from what you would expect in the financial markets.
Knowledge of these correlations is relatively transparent and can be algorithmically and programatically traded for steady, regular returns.
So how do you trade Cryptocurrencies?
Whilst it may be tempting to rely on black box trading strategies, of the sort developed by Merton and Scholes, such pureplay algorithmic trading needs to be tempered with rigorous fundamental analysis, including, but not limited to, examination of the underlying blockchain code, the status of the project development, the tension between developers as well as the application breath and community use case.
Take for instance the highly destructive Bitcoin Cash war of 2018. Underlying tensions between the various factions supporting different hard forks had been brewing for months before the hard fork had actually occurred.
Against this backdrop and with the knowledge that Bitmain, the main supporter of Bitcoin Cash (BCHABC) would need to unload its substantial stores of Bitcoin and other cryptocurrencies to fund the war to defend BCHABC against a 51% attack from the opposing camp of BCHSV or Bitcoin Cash Satoshi’s Vision — would have been an indication to the market to cash out Bitmain’s major cryptocurrency holdings — something that was made relatively transparent thanks to Bitmain’s pre-IPO documents that had been widely circulated earlier.
So what may seem like random cryptocurrency volatility is actually available for inspection well in advance of the fact. Thanks to the decentralized nature of cryptocurrencies, the transparency in Reddit channels, Twitter and other cryptocurrency forums, a keen pulse on the sentiment of the cryptosphere is invaluable when it comes to a quantitative and algorithmically driven trading strategy. Such an approach helps traders avoid black swan-type events, of the kind experienced by LTCM.
Whilst quantitative models are valuable to validate tight-band, short term trading strategies, in the long term, over reliance on these models can lead to LTCM-type disasters and extremely high drawdowns.
In that respect, a Bridgewater-type approach would probably work better (and has been proven to work better) when it comes to trading cryptocurrencies.
And because the nuances and idiosyncrasies of cryptocurrency trading generate what one trader termed a “knowledge premium,” it is possible to generate significant alpha from cryptocurrency trading while minimizing risk — just that the existing risk-management models need to be tweaked to cater for the several non-linear risks that seasoned cryptocurrency traders have lived with constantly.
For cryptocurrency traders today, mathematical rigor are not enough to generate alpha. Understanding the interrelationships between the various cryptocurrencies, the interplay within and outside the various developer communities as well as understanding the broader non-linear variables in the cryptocurrency space are instead critical for sustained and sustainable returns.
So while quantitative trading models no doubt have their value, for now at least, there’s no standard deviation to measure developer sentiment.