Systematic or model-based forecasts are challenging!

Updated: Jan 21

Written by Thomas Stämpfli, founder and CEO of Tom Capital AG

Since 1986 I am engaged in the development and commercial application of systematic forecasting models. While methods have improved dramatically over this time thanks to breakthroughs in AI (Machine Learning), results remain inconsistent. Rarely are both conditions for successful forecasting met. First, patterns from the past must repeat in the future, and second, the forecast must correctly identify and interpret these patterns.

Discovering the opportunities of systematic forecasting at Salomon Brothers

In 1986 I joined Salomon Brothers in New York. In my role as an equity salesman, I assumed accurate forecasting would become the key to adding value to clients. I looked for successful role models amongst 2’000 stock pickers with a track history in the Salomon research database. Only 40% outperformed the index in the first year; of that 40% in the first year, only 40% outperformed the index in the second year and so forth…thus even a 5-year consecutive outperformance could be coincidental. Finding the successful role model seems not easy, be it back in 1986 or today. In a recent experiment conducted by Prof. Gigerenzer, a group of German housewives beat professional Frankfurt and New York-based portfolio managers in selecting a stock portfolio. The German housewives simply selected names they knew and these in general performed better than the market. As effective as this was in the experiment, we might not feel comfortable copying it. Therefore, my resolution to focus on systematic forecasts based on algorithms that could be statistically verified seemed like a sensible direction back then and now, even with the downside that qualitative, non-quantifiable data could not be used.

My first attempt to build a systematic forecast at Salomon Brothers in 1986 was a straight success. I developed a momentum analysis model, that picked utilities to go long and to sell short in the same nominal amount. Fully hedged against market moves this strategy produced very attractive and steady returns, if transaction costs could be neglected, a privilege only a broker had in those days.

Aiming to reproduce the successes in the futures market

Together with 2 partners, we tried to replicate the success at Salomon on our own. The high transaction costs in the equity markets forced us to move to the futures markets. There we could not identify an equally homogenous group of securities required for a 1:1 replication. Therefore, we moved to technical price analysis. After two years of intensive research, we could not find statistically relevant predictors of price moves. Consequently, we gave up.

Finding a great case: forecasting credit default probabilities

In my third attempt, I did strike gold. Credit default probabilities turned out to be predictable with simple scorecards based on a few data points verified in large samples. The clever use of this led my company to obtain an 80% market share in consumer credit checks in Switzerland and Austria. Unfortunately, one of the funniest insights we could not use was the predictive strength of first names. For example, Elisabeths paid their bills with a knock-out certainty of 100% and other names did not score above 20%. It can be that easy. So, the semi-manual development of predictive scores (algorithms) has proven its worth in this field. Before too long, they beat most human credit officers because they did not have access to all the statistically relevant evidence at the time of the decision and were hampered by human emotions.

Moving back to financial markets: an intense learning experience

I am currently at my fourth attempt at generating forecasts. Transferring the insights gained in credit scoring to the financial markets (forecasts for currencies, equities, bonds, and commodity markets) yielded returns above 50% in each of the first three years. The Trump election and later the Covid19 crisis broke those patterns and forced us to adapt. There are two key challenges: The small sample size leaves the door wide open for human biases to enter, and any pattern's predictive strength is much lower than in the credit markets. There is no Elisabeth pattern around. It took us 5 years to understand and adapt to this. I will share my insight on this journey in another article titled “Why financial forecasting is so challenging for human beings”.

Tom Capital AG is an investment boutique that applies machine learning to the entire investment process, from data selection to portfolio construction. Do not hesitate to contact us at for more information or inquiries.

109 views0 comments