For five years, a group of students from Erasmus University in combination with three Robeco researchers scoured hundreds of financial newspapers from 1866 to 1926. They looked for prices, dividends and market capitalisations of almost 1500 American shares. Based on the entirely new dataset this led to, Robeco now concludes that the momentum, value and volatility (beta) investment factors also existed during those 61 years.
The research was conducted by Robeco’s Guido Baltussen, Bart van Vliet and Pim van Vliet. And from the expertise of Baltussen, who is head of factor investing at the Rotterdam fund house and is also a professor of financial markets at Erasmus University.
The outcome is a pleasant confirmation for the quant investors. “That what we do is good’, says Baltussen. “Factor investing pays off well in the long term”, he said.
That would be welcome after many factor premiums performed poorly between 2018 and 2020. “That caused a lot of questions,” explained Baltussen. “Many clients started at the peak, in 2017, 2018. The fact is that, since then, it is precisely the tech giants that have done well, while the rest of the market has lagged far behind.”
Tech companies are large, expensive and often volatile, which is why they are not likely to be in the portfolio of a factor investor. According to Baltussen, this is the reason why factor funds have lagged behind. Until last year, when value and momentum stocks started to perform well again.
Concerns about p-hacking
However, the reason for the study published in November was not so much uncertainty about the strength of factor premiums. Baltussen explained his involvement: “On the one hand, it was my role as a professor that led me to do this research, because p-hacking (discovering patterns that appear to be statistically significant, when in fact there is no effect, ed.) is a topic there. Academically, it’s a concern as to whether this is actually happening, and in which areas.”
But also in his role as head of factor investing at Robeco, Baltussen felt it was important to start testing with a new dataset, because factor investing is after all widely used, for example by Dutch pension funds. “There are plenty of studies on factor premiums, which show the effect of factors such as momentum, volatility and profitability. The only problem is that most of these studies are based on a sample of American shares over a period of fifty to sixty years - usually after 1963.”
“This dataset called CRSP has been completely wrung out,” said Baltussen. “What new insights will it provide and how reliable is it? If you test an idea with 95 per cent reliability, there is a 5 per cent chance that you think you have found something but it is not. If twenty people test the same idea on the same dataset, the chance suddenly becomes very large. The CRSP dataset has so many users that we are not talking about twenty people, but -say- twenty thousand. That could mean that you have to take the practical investigations on that dataset with a larger grain of salt, that they may be due to a coincidence in the data.”
One way of dealing with this is to do out-of-sample tests, i.e. to use other datasets. Robeco uses these too, datasets in Europe and emerging markets. But according to the researchers, these sets have also been tested.
That triggered them: what was a dataset that had not yet been tested and would be relevant for Robeco as a factor investor? The researchers wanted to construct a new database of an economically important period for America, with strong economic growth and rapid industrial development.
Five years of construction
They decided to go for the period 1866-1926, one in which the stock market was comparable to today’s market in many ways.: “Good liquidity, many experienced investors active”, Baltussen pointed out. “Factor premiums come from people’s behaviour. Behaviour is something that evolves over millions of years, but not over hundreds of years. So you would also expect to see momentum, value, low risk reflected in the behaviour of the investors of the day.”
The Rotterdam Erasmus University and Robeco built a database, after which they started validating the data and values of the capitalisations. “A job that took about five years. In the late 1800s, many bank shares were floated on the stock exchange. There was a bank in every small town, comparable to the number of bakeries today. Economically irrelevant, from a practical point of view you didn’t invest in them.”
But with equal-weighted data, you do end up with all those kinds of names in your portfolio. That is why the researchers decided to develop a market-weighted dataset, so that the value of a company would reflect its value in the economy. “As far as I know,” said Baltussen, “we were the first to do this for the data over this period.
Size no premium
An enormous gathering and overtyping job for a group of students, followed by the research by the Robeco researchers. The factors momentum, value and volatility emerged as hoped and expected as robust factor premiums. Size did not. ”But that is something that has been becoming increasingly clear for some time,” said Baltussen. “Size seems more like an amplifier of other factors, not a stand-alone factor premium.”
So, more generally, it can be said that what other researchers have found in the existing CRSP dataset has now also been found in the dataset for an earlier period. And that these studies therefore still make sense, although remaining critical is the motto according to Baltussen.
The “pre-CRSP” dataset will become available for other researchers, so that more research can be done about this period before 1926.