Investor Decisions Part 1: Applying Machine Learning and Science to Trading Decisions

TL;DR: Treat investments like hypotheses to be tested not as risk adjusted returns to be maximized.

The adage: ‘cut losers and let winners ride’ needs revision (to put it kindly), and if you use stop-loss-levels, you should consider using stop-loss-curves.

This revisits the content from an ‘oldie-but-goodie‘ talk I gave in 2016. I was still breaking out of my Hedge Fund / Portfolio Manager comfort zone and was starting to help the Boston Machine Learning community and the Boston Asset Management community learn from each other’s relative strengths.

"No man is better than a machine… and no machine is better than a man with a machine.”

Paul Tudor Jones – 2016 (Forbes & WSJ )

This was an exciting set of topics for me because I began formalizing how the various insights of machine learning could lead to enormous process efficiencies for investment managers. I began formalizing what insights active management bring to the table and that they are fundamentally different than their silicone-based forecasting counterparts. Some investors like Paul Tudor Jones stated: ” No man is better than a machine… and no machine is better than a man with a machine.” This has the right sentiment but it isn’t very specific as to why there are these synergies.

An Avocado Anecdote

The reality of machine learning or any type of statistical analysis is that it is backwards looking. The engineers or investors building the quantitative models inevitably have to rely on historical relationships to tease out patterns and most of the time this works extremely well, especially when forecasting highly specific things. For example, data could answer how La Nina could inflate avocado prices and impact Calavo’s quarterly revenue numbers, but it couldn’t answer the question about how it would affect the margins and profitability of Calavo that quarter (1Q2017). Using weekly avocado prices we could see the slowing of sales and the 40% spike in prices. Use quarterly balance sheets to estimate company margins and changes in input costs, the confidence bands on the forecasted margins for the company remained wide.

View Original Tweet

We concluded in our report that “Supply, on the other hand, is having problems keeping up. An unexpected cold snap this winter led to a worldwide collapse in avocado production. California avocado farms suffered the worst, dropping in production volume by almost 50% year/year according to data released by the Hass Avocado Board. Mexico followed close behind with a 20% year/year decline in production quantity. Price elasticities suggest that, even with the demand spike, avocado revenues in the US could be down around 18%.”

The Guacamole Situation

There are limitations with this type of analysis. We only had limited histories of company data to review and it was rare for avocado prices to spike this much. To top it off, the spike occurred just before Cinco de Mayo! The quantitative models don’t understand how inelastic the demand for guacamole is at that time of year and it may not have a sense for how Calavo could pass along the prices to consumers and distributors.

  • If you believed the historical demand patterns and elasticity for avocados would continue, then the earnings-per-share (EPS) for Calavo $CVGW would have disappointed and likely the stock would have declined.
  • If you believed that due to the holiday or perhaps due to the emerging love of avocado toast among my peers that demand for avocados and that ‘this time is different’, then EPS for avocados would surprise to the upside and the stock should rise.

Ultimately we decided to go with the models and assume a more constant elasticity in the demand for avocados. Our revenue forecasts were dead on. Our EPS forecast… the one that primarily drives stock moves on earnings day, was too low. Demand was strong for avocados, EPS exceeded expectations and the stock went up. C’est la vie

This shows the essential conflict between extrapolating historical relationships and overreacting to perceived changing market conditions. Although in this case, sticking directly with the data and underweighting this special situation led to a worse forecast, this is not always the case. Human’s make errors and this wikipedia article on Cognitive Biases is very long.

Striking the balance between using data and algorithms to identify emerging trends and give you confidence in the persistence of known relationships, while relying on humans to shift the types of inputs and to question whether ‘this time is different’ seemed to be the key way to integrate human insight with a machine’s probabilities. These days I simply state that:

“humans give you possibilities, Machines give you probabilities”