What is the future of Algorithmic Trading? by Laurent Bernut
Answer by Laurent Bernut:
If investment is a process, then automation is the natural conclusion. “He, who controls the past controls the future”, George Orwell. Before attempting to guess and inevitably fail what the future holds in store for us, let’s take a step back and understand why they are not so prevalent today
Complexity is a form of laziness and the gaze heuristic
Investors inherently distrust “black boxes”. We can accept someone giving us bad directions. We hate when Google maps sends us off course. This has implications on AUM.
Computing power is not the answer, it is merely one solution. AI is an attempt at solving complexity with complexity. Complex problems can be solved with simple solutions. Example: a fast moving projectile bouncing around is rocket science, literally. Now, when was the last time Serena Williams solved stochastic equations? Gaze heuristic. The point is that behind every algo there is a person formalising her current perception, fears, beliefs of the markets.
The hard part is articulating complexity in simple heuristics and only then apply sophisticated solutions. Algorithms that survive have to be robust. Complexity is fragile. For that reason only, i do not share the current enthusiasm on AI. We live in a technological world, governed by sophisticated algorithms (train schedules, traffic lights, chromatography etc). Yet, we have failed to apply our genius to the markets. AI is just the latest attempt.
Instead of thinking about the model logically, we try to discover it through tuning, optimisation, Monte Carlo, neural networks, AI, whatever. It is exactly like tuning an engine until You realise You need to build something called a carburetor.
Asset Allocation algorithms
To me, the immediate future of algorithms starts with asset allocation. There is a growing consensus that asset allocation is the primary driver of performance. Robo-advisors are still in their infancy, but this is where the industry is heading.
Unfortunately, they rely on Markowitz CAPM. It was invented in the year that the US Army drafted Elvis. Since then, doctors have stopped prescribing cigarettes to pregnant women, people have walked on the moon and non-alcoholic beer was invented
Here are a few things that Markowitz did not think of:
- Bear markets: this has been a 30 year mega bull markets for bonds. We are officially in the longest equity bull run in history. What happens when trends revert? Correlation will go to 1. Short selling is the asset class that remains terra incognita for the overwhelming majority of investors
- Time (synchronicity) is a form of asset allocation: mean reversion is a time frame issue. Any security trends on longer and/or shorter time frames. When trading the same strategy on the same security across multiple time frames, synchronicity smoothes the equity curve
- Volatility is NOT a measure of quality: Markowitz defaults to volatility as proxy for quality. CDOs, lending club and sovereign bonds are quality until… Integral of drawdowns is a robust measure of quality
- Position sizing is a form of asset allocation: this is where algorithms will find their first low hanging fruit. Risk parity has correlation issues
Time is the wrong container: 1st derivative of time, price and volume is speed
HFT is here to stay. Whether they should be legislated and taxed away is a different debate. For now, most algos fail because they are conceived on wall clock time. Algos will gravitate toward HFT when they are conceived with speed as opposed to wall clock time in mind.
Flash crashes happen in thin markets: mid-afternoon, middle of the night. This is because they are designed to react to wall clock time. We have been conditioned to see time as linear, wall clock time. Volume is not distributed uniformly through time.
Picture a bottling line. Bottles move only when they are filled. Now, imagine two shifts, day shift with everyone along the line, the other with two people and bottles moving every 5 seconds regardless. When everyone is here, they would overflow. When no-one is there, they would move half empty. It does not make sense, does it ?Time is the wrong container: 1,000 or 20 ticks per second are different markets.
The first derivative of time, volume and price is speed. HFT algos do not need to be so quick to react in slow moving markets. Conversely, conventional algos will gravitate toward higher frequency when they are set on speed (i-e constant volume) as opposed to time. Time is the wrong container.
Short selling algorithms
Contrary to popular belief, short selling is the province of algorithms. Firstly, successful shorts shrink, successful longs balloon. So, there is a need for more ideas, more often just to hedge exposures. The way to accomplish this is to systematise the process, hence algos
Secondly, on the short side, the market does not cooperate. Write and test algos for the short side. If they survive the short side, they will thrive on the long side.
Thirdly, the secret to raising AUM is to master the short side. Ask John Paulson about this. Next bear market, when everything drops 50%, all this AI, factor optimisation covariance matrices mumbo jumbo is highly unlikely to garner any sympathy from disgruntled investors. You may not like it, but if You intend to survive, do You have a choice?
The privilege of simplicity is that it imposes itself, even to those who do not understand its sophistication.
My own journey into the algorithmic world has been an unquenchable thirst for simplicity. When faced with a complex problem, I work until a simple solution emerges (the answer is rarely better signal processing by the way). This is like peeling an onion: solving one problem leads to the next one. In the end, algos are like iOS, next update is coming soon