How to Build a Dynamic ETF Portfolio using Artificial Intelligence

October 18, 2017

Building an ETF Portfolio using Artificial Intelligence

In the following, we present how to construct a high-performing dynamic ETF portfolio based on the algorithmic forecasts generated by our artificial intelligence system which including the costs of bid-ask spreads and commissions, results in:

Returns of up to 40% in a 2-year time period
Alphas reaching 18%
Sharpe ratios reaching 1.2

We first present the performance of trading long and short only the top ETFs as identified by our algorithm’s ranking system and then evaluate how these results can be combined with a market baseline into a long only strategy that allows investors to target desired portfolio statistics.

About I Know First

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. We generate investment signals for a universe of over 3.000 assets which result in a daily ranking of investment opportunities. Our algorithmic forecasts can easily be integrated into the daily 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 an effective ETF portfolio using the signals generated for the 11 SPDR sector ETFs: these are ETFs of S&P 500 stocks grouped by GICS sector classification which facilitate passive investment in specific sectors of the US economy.

ETF Portfolio of Top Sector(s)

First we present the performance of constructing an ETF portfolio by investing long and short in the strongest two sector ETFs as selected by our algorithm. The results of this strategy including the effects of spreads and commissions can be seen in the tables below. The first table shows the results on a portfolio level and the second table on a trade level (click on the tables to enlarge).

The first row in the two tables presents the statistics of investing only in the ETF with the strongest signal, while the second row of investing in the top two ETFs.

As can be seen above the two ETF portfolios have an outstanding performance:

Total returns reaching 40% and 33% versus the benchmark’s return of 22% in the same time period
Sharpe ratios above 1.00 versus the benchmark’s 0.81
Betas of -0.01 and 0.12
Alphas above 14%
Mean returns per trade over 0.25% with holding periods of ca. 4 days, resulting in annualized returns of 14% and 18%

The chart for the two strategies can be seen below.

The ETF portfolios result in steady and consistent growth over the benchmark even in this period of great market growth.

ETF Portfolio of All Sectors

In a second step, we combine these results with a baseline equally invested in all sectors to construct a long-only ETF portfolio which follows the market but overweights/underweights sectors according to our algorithmic predictions. This is implemented by initially investing equally in each SPDR sector ETF and then adding weight to those sectors for which our algorithmic signals are positive and subtracting weight from those sectors for which the signals are negative. This results in strategies that are long only as the minimum weight for any sector is zero and which can trade off proximity to the overall market for alpha by moving weight from the baseline to the algorithmic signals. The statistics of these strategies are displayed in the table below.

The table can be read from the bottom up as going from investing equally long in all SPDR sector ETFs without using the algorithmic signals (benchmark, row 5), gradually adding weight to I Know First’s predictions resulting in a combination of the benchmark and the algorithmic forecasts (rows 2-4), and culminating in the top row in an ETF portfolio in which only the algorithmic signals are used (row 1).

The bottom two rows in the table are the two benchmarks: equally weighted SPDR sector ETFs (row 5) and fully invested in the SPY (row 6).

Rows 2-4 show the combination of the algorithmic signals and the equally weighted SPDR sector ETF benchmark, as explained above. Row 4 gives equal weight to the benchmark and the algorithmic signals, while row 2 gives five times as much weight to the algorithmic signals as to the benchmark.

Row 1 shows the result of investing in only the sectors that have positive algorithmic forecasts with a maximum of 25% of the portfolio allocated to any sector.

As can be seen in the chart as we move from row 5 up and thus go from an equally weighted ETF portfolio to a more algorithmically weighted ETF portfolio:

Total return increases from 17% to 28% with all I Know First portfolios outperforming the SPY’s 22% return
Alpha increases from 0% to 5%
Beta decreases from 1.00 to 0.72
Sharpe ratio increases from 0.79 to 1.20
Volatility remains stable at around 11%-12%.

This approach allows investors to fine tune their ETF portfolio to the statistics and proximity to the benchmark they desire by allocating more or less weight to the algorithmic signals with the best performing strategy being the one that only invests in the ETFs selected by the algorithm.

The equity lines for the various ETF portfolios are displayed below.

The chart displays the same pattern described above and shows the strategies give rise to smooth, consistently growing equity lines.

Conclusion

In closing, we presented a set of analyses on how to construct an ETF portfolio using the signals generated by our self-learning predictive AI algorithm for the SPDR sector ETFs.

First we showed the results of investing long and short in the top 2 sectors selected by our algorithm which resulted in portfolios with:

Total returns reaching 40% in a 2-year time horizon versus the benchmark’s 22%
Sharpe ratios above 1.00 versus the benchmark’s 0.81
Mean returns per trade over 0.25%

We then combined our algorithmic forecast with the benchmark to construct long only strategies that base themselves on the benchmark while generating alpha by overweighting and underweighting sectors according to the algorithmic signals. This analysis clearly shows that increasing the weight of algorithmic predictions results in a better performing ETF portfolio with: