ETF Predictions Based Trading Strategies Using I Know First’s Aggregated ETF Forecast

We continue our series of articles analyzing the performance of ETF predictions computed by aggregating I Know First’s algorithmic forecasts for the individual stocks contained within the funds. In this article, a new method to generate an ETF forecast by using a weighted aggregation of stock-level predictions is presented and the performance of this method is tested using Sector ETFs. We show that the computed ETF predictions result in strategies which outperform the benchmark and present excellent performance and risk statistics both on an individual ETF level and for portfolios of ETFs.

I Know First Overview

I Know First is an Israeli Fintech company that brings artificial intelligence to the financial world by providing daily investment forecasts based on an advanced self-learning algorithm. This algorithm generates investment predictions for a universe of over 8200 assets which result in a daily ranking of investment opportunities. These can easily be integrated into investment selection processes and, combined with the appropriate strategy, be translated into portfolios with outstanding statistics for all types of investors.

Here we focus on ways of constructing ETF predictions from weighted individual stock level predictions. We then use these types of investment signals to build trading strategies for the 9 SPDR Sector ETFs and the Vanguard Real Estate Sector ETF individually and as a portfolio of ETFs. These investment funds contain S&P 500 stocks grouped by GICS sector classification and facilitate passive exposure to specific sectors of the US economy.

ETF Forecast Using Weighted Stock-Level Forecast Aggregation

The general concept behind this ETF forecast is to combine the information from individual, stock-level forecasts for the assets contained within the ETF to decide whether we are bullish, bearish, or neutral for the entire fund. For the method presented in this article we take the components and weights of the stocks the ETF is made up of as stated on the ETF provider’s website and map them to our daily algorithmic stock forecasts. Thus, we sum the weights for which our predictions are long and those for which they are short and divide by the total sum of the weights, which gives the weighted percentage of stocks within the ETF in each direction. Finally, we use a fixed threshold level to compute the final ETF forecast: if neither percentage passes the threshold we set the ETF forecast to neutral. This process results in a daily investment decision the performance of which is shown below through backtests of investment strategies.

Individual ETF Predictions’ Performance

We first show the performance of using the aggregated ETF forecast for each ETF individually. Each day we compute the ETF prediction for the specific day and invest in the direction selected. The following table shows the results of trading using this method (without including commission costs) versus going long the ETF for the period 08/18/2015 – 11/01/2017 for the 9 SPDR Sector ETFs and the Vanguard Real Estate ETF.

As can be seen in the above table the ETF forecasts outperform their respective benchmark 7 out of 10 times in terms of risk adjusted return (Sharpe Ratio) and on average outperform the benchmarks in terms of Total Return (30% vs 27%) and Sharpe Ratio (1.00 vs 0.8) with average Alpha and Beta of respectively 11% and 0.2.

Below we show the equity lines for the various ETF forecast based strategies (in green) versus their benchmark (in red, click on the images to enlarge).

 

The equity lines show how the algorithm de-risks and avoids market downturns while also signaling to open positions in the bull markets.

Performance of an ETF Portfolio Based on the Aggregated Predictions

In a second step, we combine the ETF predictions for all 10 ETFs included in the previous table to create an ETF portfolio which invests equally in each ETF in the direction given by the aggregated ETF predictions.

The following table gives the performance of this strategy for a long/short and long-only portfolio (where all short predictions are treated as neutral) restricted to allocate at most 25% of the portfolio value to any individual ETF (i.e. if we need at least 4 non-neutral positions to be fully invested). In this case, the SPY is used as benchmark (ETF that tracks the performance of the S&P 500 index).

The portfolios present excellent portfolio statistics with Total Returns of 46% and 38% over the 2-year period, strongly outperforming the benchmark’s 28%, Sharpe Ratios above 1.5 versus the benchmark’s 0.99, and annualized alpha of 19% and 10%.

Below we show the equity lines for the two strategies (blue and green) against the benchmark (in red).

Conclusion

In this article, we presented results of using ETF predictions computed through the weighted aggregation of I Know First’s algorithmic stock predictions of the stocks which compose the ETF. Investment strategies based on these forecasts outperform their relative benchmarks on both an individual ETF level and for a portfolio of ETFs with the final ETF portfolios constructed using the weighted aggregation over a two-year period yielding:

  • Total Returns of 46% and 38% versus the benchmark’s 28%
  • Sharpe Ratios of 1.65 and 1.86 versus the benchmark’s 0.99
  • Annualized Alpha of 19% and 10%