Portfolio risk forecast

In practice, portfolio managers use either the fundamental model or the statistical model, or a combination both, to model the volatility of a portfolio’s future performance.

• The fundamental model specifies characteristics that influence the covariances of equity returns. For example, stocks in the same industry will have more correlated returns with each other than stocks in different industries. Financial leverage or the size of firms also seem to explain part of the common variation in returns. Using a list of these characteristics and their time series as well as historical stock returns, it is possible to estimate the covariance matrix of risk factor returns in order to calculate a security’s exposure to these factors. In addition, by aggregating them we are able to calculate the risk of any portfolio.
• The statistical model starts from a full covariance matrix. This has often been calculated with returns filtered by a “winsorization” process. A principal component analysis (PCA) identifies the main dimensions of the covariance matrix by ranking them by their importance but without identifying the factors to which they refer. The risk factors are inferred empirically.

Until now GPTD has built risk models according to these two types of modeling. Statistical models (which use less information) work well, especially when stock markets are less volatile but are less adequate when market volatility increases sharply. So far, little consensus exists on the optimal way to combine the fundamental model or the statistical model or on the possibility of constructing a third one. TDAM is looking at what would be the best way to forecast the risk of a portfolio that uses historical returns on stocks in a certain universe and the fundamentals of those stocks?