Minimizing portfolio drawdown for long only strategy, market crash of 2018.
Updated: Nov 6, 2019
In the post we present backtesting results for minimizing drawdown of a portfolio of stocks during the market crash of late 2018. Drawdown is a maximum decline of a stock/portfolio value from its peak point during the investment period, see Figure 1. Although the stock achieved ~15% return in period of 100 days there was a large drawdown of ~20%. Such drawdown may not be acceptable for some investors. Our software builds a portfolio which minimizes drawdowns during the investment period. Moreover, you can choose pessimistic macro-scenarios in the models, e.g. stong volatility, autocorrelations, recesssions, to achieve more robust optimization results.
We use the same list of 300+ stocks, randomly picked from the US, European and Asian markets as in previous reports. The training data contains 1000 daily return observations until January 2018. The testing data contains ~300 daily returns observations between January 2018 and January 2019. Leverage is 1 and investment into a single stock is limited to 5% of the total investment. In this case we apply enhanced macro-scenario modelling assuming that the market should not grow in the coming ~100 days. Hereby, we would like to minimize the drawdown of our portfolio specifically for flat/moderately bearish markets.
The solution from the frontier is selected using our risk preferences using the frontier tangent logic, i.e. when incremental expected return is lower than the incremental increase in drawdown. First, we demonstrate performance of the model on training stock data, 1000 return observations before January 2018 (Figure 3). We can observe that there are only very small drawdowns of the portfolio based on training data with returns ~2x higher in comparison to benchmark.
Next, we check the out-of-sample performance of the model in 2018. The benchmark represents the average return of all stocks in the short list.
The drawdown of the optimized portfolio is significantly lower than that of the benchmark, ~-10% vs ~-17% respectively. Overall, the optimized portfolio strongly outperforms the benchmark ~+5% vs ~-5% respectively, total for the year. Drawdown models are much more computationally expensive and take more time and resources to implement. We used only 200 sample paths with 100 days returns for the model which is obviously not enough when had a selection of 200+ stocks. Nevertheless, we confirm significant potential of the models for risk reductions, especially if applied together with enhanced macro-modelling.