Taking a brief respite from my usual topics of startups, investing and personal development — today, I wanted to touch on finance & science. You may or may not know that last year I joined AIG as Global Head of Product within Science. Science is a machine learning focused team responsible for pioneering algorithms that both mitigate risk and generate returns within equity and fixed income markets, assisting in the management of over $320 billion.
Anyway, back to the post — in January a paper titled “Global Risk Premium” was published, in which a group of quants from Robeco Groep, the Dutch investment house, produced research into the global performance of stocks, bonds, currencies and commodities going back to 1800.
One of the authors recently introduced their thinking on LinkedIn:
“There is a crisis in science. Too many studies cannot be replicated. There is bias to positive results which is referred to as ‘p-hacking’. This is caused by incentives to publish positive results. If you can easily test 20 hypotheses, one will turn out to be significant. “Torture the data until they confess’ they say. This p-hacking is a serious threat to science itself, but it can be addressed. For example, by raising the statistical bar. Furthermore, replication studies have become more common in social sciences. The Economist wrote about this issue in 2013 and Campbell Harvey has put p-hacking on the financial research agenda.
At Robeco Quantitative Investing we manage billions based on scientifically proven investment factors. These factors need to have strong evidence and need to have a clear economic rationale. Nobody knows 100% sure if these factors keep on working in the future. But what if the scientific results were fake to start with? We therefore apply the same cures which are proposed by the leading scientists. Replicate previous studies and raise the statistical bar. As conservative fund managers it is our fiduciary duty to apply the highest possible standards when making investment decisions.
Over the past years Guido Baltussen, Laurens Swinkels and I have constructed a cool deep historical dataset stretching back to 1800 and test 24 global factor premiums. We replicate several previous studies which typically go back ‘only’ 30-40 years. We have some remarkable findings. Full article can be found on SSRN, a platform for social science research”
With the above introduction in mind, and having read the paper, there are a few conclusions that I have drawn — feel free to tweet me @ScottTaylor with your conclusions!
- Correlations between factors and between asset classes are remarkably constant over time
- Over the two centuries, trend-following worked better and more reliably than any other factor, with a strong and consistent risk-adjusted return
- Seasonality is one of the most reliable investment factors
- Behavioural finance vs. market efficiency – anomalies do appear to be real, and the market is not as efficient as it is supposed to be. In fact, it gets some things systematically and predictably wrong