Non-parametric forecasting for conditional asset allocation

Abstract

Conditional asset allocation (CAA) involves using key past economic and financial data to produce forecasts of expected returns for the various asset classes involved in the asset allocation decision. Traditional forecasting models for asset returns, in particular linear regression models and ARIMA time series based models, often provide economically meaningful asset allocation decisions. These results, however, only seem to possess power in-sample, and the out-of-sample performance of such methodologies suggests, consistent with efficient market theory, that there is no economic benefit from undertaking conditional asset allocation.

Alternative forecasting models have only rarely been explored in the literature. Here, the economic forecasting power of kernel regression, a fairly common non-parametric statistical model, is investigated. As for the more traditional methodologies, it is found that the model has good economic significance in-sample. Unlike in the more traditional methods, this economic significance also persists in out-of-sample periods. Moreover, simple trading rules based on filtered signals from the model provide enhanced economic performance. The reasons why this alternative methodology provides better out-of-sample results are conjectured, and it is tied to the behaviour of the underlying asset prices.