A machine learning algorithm creates an asset return forecast based on market, fundamental, macroeconomic, and text data.
In backtests, combining this forecast with lagged average returns leads to the best performing strategy.
The asset class historical returns data is used to estimate the risk of the different asset classes.
The forecasted asset class returns and anticipated risk measures (based on historical data) are fed into a portfolio engine which generates the portfolio with the highest possible return at each level of risk.
For tax efficient versions of our strategy, the portfolio engine will ensure that its selected portfolio does not deviate too much from last month’s optimal portfolio.
The lower risk portfolios have a higher allocation to bonds and the higher risk portfolios allocate more to equities.
Leverage is not allowed so all the portfolio weights (i.e., the fraction allocated to each asset class) are between zero and one.
Human judgment is critical at two points in the process: (1) the choice of the precise amount of risk at your level of risk tolerance, based on an assessment of market and economic conditions; and (2) vetting the model-generated portfolio to ensure the model isn’t missing first-order information about current market and economic conditions.
Each asset class may potentially be predicted by numerous forecasting variables.
Which variables are used to forecast and how they forecast changes over time. We reduce a long list of forecasters to a much shorter list; this is governed by a smoothness parameter. We decide over what time frame to estimate the model mapping forecasters to future outcomes; we refer to this as the training horizon.
For a given choice of smoothness parameter and training horizon, we use rolling windows over the data to estimate the relationship between asset class returns and forecasting variables. We use a machine learning technique called a lasso (least absolute shrinkage and selection operator). This generates out-of-sample (not using forward information) forecasts, which we use to assess the quality of the model’s return forecasts.
We then choose a smoothness parameter and training horizon to optimize the model’s out-of-sample forecasting performance.
The lasso model generates a next 12-month return forecast for each asset class in every month. The set of forecasting variables changes over time, reflecting changing market dynamics.
We begin with our proprietary machine learning based asset-class return forecasts from the prior step.
We combine these forecasts with dynamic estimates of risk and correlation for the asset classes to form the highest possible forecasted return portfolio for a given level of targeted risk.
For tax efficient versions of our strategy, we ensure that the highest expected return portfolio at a given risk target does not deviate too far from last month’s optimal portfolio. This allows our investment strategy to respond to market conditions, and respect clients’ risk targets, while avoiding excessive trading and excessive realized gains (though some realized gains are inevitable).
This step requires that all asset class weights are non-negative (to rule out short-selling) and that each asset-class allocation does not exceed an allowable threshold (for example, we do not want to allocate half of the portfolio to REITs).
An important parameter in this process is the time horizon over which the risk and correlation estimates are made. Too short a horizon introduces too much noise; too long a horizon slows response to changing market conditions.
In portfolio simulations, we found that using average returns, as opposed to the model-based return forecasts, also led to good portfolio performance.
The optimal strategy combines our model-based return forecasts with the mean return of each asset class calculated over the time horizon used for the risk and correlation estimates.
The two important parameters in the portfolio construction step are:
We searched over many possible combinations of these parameters to find the historically optimal pair.
An important input into the portfolio engine is the asset class correlation. This measures the tendency of asset classes to comove. Here we see the asset-class correlation matrixes estimated over the prior 1- and 4-year periods.
Negative or low correlations between asset class returns generates diversification benefits when allocating across asset classes.