How We Use Self-Learning Algorithms
The Wizard of Odds was a master of probability.
In a break from our usual long-form pieces covering businesses and valuation, this edition focuses on how we use self-learning algorithms in our internal framework.
Our basic philosophy leans away from predicting future outcomes.
We are better served by triangulating embedded market expectations, then corroborating/diverging with embedded expectations. Our self-learning algorithmic framework is built on a similar game theory philosophy.
Specifically, our self-learning algorithmic framework is built on the following foundations.
- Instead of imposing a user-supplied, hard-coded version of reality (e.g. ‘Over 40x P/E is expensive’, or ‘we will only use the 50/200-day moving average’), the framework needs to learn the language spoken by each financial instrument. As it is rare for two instruments to behave in the exact same fashion, a humble self-learning system ought to learn from individual instruments. More importantly, it must learn to dynamically change its learning model, as the underlying instrument’s behaviour evolves with time.
- It must maximize overall outcome probability, on a risk-adjusted basis. Most systems look to maximize performance. A more prudent approach from a probabilistic survival angle is to adopt a risk-optimized philosophy. The system should have positive survival probability over time.
- It must have a flywheel effect. Once a learning system is built, it must serve as a driver variable in other parts of the investment value chain.
- It must operate at scale, and be instrument and geography independent. This allows us to seamlessly apply the framework across global financial markets and instruments.
- It must optimize for relative edge. Most investment strategies suffer from alpha decay: as a strategy attracts more participation, alpha generation tends to decrease with time. The self-learning system needs to be in listen-only mode as default. It should absorb market information, learn to avoid overpopulated markets/strategies, and seek out underpopulated markets/strategies to maximize relative edge over market participants.
This framework is best described through an example of our self-learning Volatility strategy (the ‘BollSig Algorithm’).
BollSig Algorithm Signal Triggers
- Look-back period.
- Volatility.
- Optimizer Layer.
The BollSig Algorithm triangulates instruments based on optimized Signal Triggers. (1) and (2) optimize for history and price volatility, (3) does the hard job of dynamically absorbing and risk-optimization of regime switches (alterations in the behaviour of the instrument over time). The strategy assumes no Short positions.
Example: EUR/USD
The Algorithm is illustrated using the EUR/USD FX pair.
The Algorithm signals a Buy at Green markers on the price chart; and unwinds at the subsequent Red markers. Pink line (RHS on chart) is the Instrument Price. Cyan line (LHS) depicts the Cumulative Return on this strategy.
Quantitative Algorithms tend to have hard-coded triggers. This inhibits strategy functioning, as the user provided hard-coded triggers may not be optimal for the instrument under consideration. A more dynamic self-learning mechanism is a step improvement.
This self-learning element has been embedded into the BollSig Algorithm through a process that identifies the best Signal Trigger combination for each instrument under consideration.
This enhancement ensures that:
Optimal Signal Triggers are captured for each instrument at any point in time.
As instrument behaviour evolves over time, the system dynamically alters the Signal Triggers to pick the most optimal combination.
Selected depictions of various BollSig optimization runs for EUR/USD are shown. Cyan = Best combination. Orange = Worst combination.
A Buy-n-Hold strategy covering the full period would have yielded an annual return of +1.13%.
As the various runs depict, some combinations lead to negative returns as well. A self-learning approach — as opposed to a user-supplied, hard-coded Signal Trigger approach — guards against costly experiments.
The evolution of the self-learning approach over time is illustrated below for the EUR/USD pair for various randomly sliced time periods.
The optimal combination is highlighted for each subperiod. For comparison, various Signal Trigger combination performance is contrasted with a simple Buy-n-Hold strategy.
As the behaviour of the reference instrument evolves with time, BollSig dynamically evolves to the most optimal combination.
In many cases, BollSig generates an Exclusion Signal — the instrument in question is not an optimal choice for this strategy. This Exclusion Signal is valuable, as it bypasses instruments with suboptimal, long-term risk-optimized performance.
Flywheel Effect
Once a learning system is built, it must serve as a driver variable in other parts of the investment value chain.
BollSig is used:
- On a standalone basis.
- As an overlay on existing long-term investment holdings.
- For Implied Volatility opportunities.
An instance of how we utilize the Flywheel Effect is shown through (3).
BollSig also helps triangulate the prevailing Implied Volatility Term Structure (distribution of Volatility expectations across various time periods). Currently, EUR/USD is reasonably cheap in terms of Implied Volatility relative to its own history. This serves as a foundation for Volatility capture investment opportunities.
Static quantitative systems suffer from various constraints. One of the most difficult nuances to grapple with is Behavior Drift: alteration in behaviour of the underlying instrument.
As we show here, a self-learning algorithmic framework offers a several fold improvement over a static/hard-coded approach.
The system’s versatility lies in its ability to do this:
- At scale.
- Customized for each instrument of interest.
- Across markets, and geographies.
- Through the Flywheel Effect; by feeding into other parts of the investment value chain.
