Events

Backtest overfitting means the usage of backtests (historical market data) to construct an investment strategy, fund or portfolio, when the number of variations explored exceeds limits of statistical reliability. We show that backtest overfitting is inevitable when computer programs are employed to explore millions or even billions of parameter variations (as is typical) to select an optimal variant. We illustrate this by showing that while it is a simple matter to design a stock fund, based only on a weighted collection of S&P500 stocks, that achieves any desired performance profile,...

We begin with a survey of machine learning techniques and applications outside of finance. Then we discuss our use of Machine Learning techniques at Rosenberg. Finally, we explore some alternative data sources.

We formulate the portfolio construction problem as a mean/variance problem which includes a linear term representing an investor’s preference for expected “social return”, in addition to her expected “financial return” of the classical theory. By making various assumptions, we are able to exploit the heterogeneous expectations version of the CAPM to derive an equilibrium model which is an extension of the standard Capital Asset Pricing Model. Among other things, the model implies that, in equilibrium, assets with higher expected social return that is valued by investors will have, ceteris...

I examine whether whether low capital levels incentivize banks to systematically originate and hold riskier loans. I construct a novel data set consisting of 1.8 million small business and home mortgage loans, matched to the specific banks that originated them and the capital levels of those banks at the time of origination, and verified to be held on bank portfolios, rather than sold. A one point increase in capital ratios (e.g. from 12% to 13%) is associated with a 4% decrease in the default risk of mortgage loans held on portfolio (from a net foreclosure rate of 2.5% to 2.4%). Bank...