This article was co-written by Ellie Lan, an Investment Analyst at Betterment, and Dan Egan, Betterment’s Director of Behavioral Finance and Investing.

When you use algorithms and data to make decisions, you agree to a very basic principle: You clearly and logically lay out the steps you follow, and then use the answers that your analysis produces—whether or not that jibes with a personal opinion or feeling, or worse, incentive provided by someone other than your customers.

At Betterment, this is a philosophy we apply at every level of the company—including how we selected the funds that are used in our portfolio.

As mentioned in our portfolio selection white paper, the following core criteria drove our ETF selection process:

  • Positive expected return, after adjusting for taxes and risk
  • Low cost
  • Highly liquid
  • Passive/index-tracking
  • Low turnover
  • Tax efficiency

We considered the full universe of U.S.-listed exchange-traded funds (ETFs). For our ETF selection, we created a filtering and ranking algorithm (using the free open-source programming language R) to choose ETFs that met our core criteria and ranked highly on qualities aligned to our investment objectives. We applied these filters to remove asset classes and investment vehicles that did not meet our standards, and then selected from the top remaining candidates.

We assessed the universe of 1,654 ETFs (as of Sept. 4, 2015). This transparent process resulted in 30 different ETFs for customers to invest across 13 asset classes. We outline the type of ETFs we sought as well as the ones we actively avoided. The process reduced average expense ratio for the overall set from 0.64% to 0.20% for the final set.

ETF Exclusion By Category


Asset Class

An asset class is a group of investments that share similar characteristics and behaviors. Funds with an asset class of equity or fixed income were included in our analysis. Funds that were in alternatives, commodities, asset allocation, and currency categories do not align with our investment objectives. They are either highly speculative, expensive, or generate uncertain expected returns.

The average expense ratio of ETFs that were filtered out was 0.97%, while that of the included set was 0.57%. The spread was similarly reduced from 0.15% to 0.07%. By filtering out asset classes with high expense ratios and bid-ask spreads, our algorithms can steer clients away from expensive and illiquid investments right from the beginning.

Inverse ETFs

Inverse ETFs are constructed using derivatives that allow the investor to profit from a decline in the value of the underlying benchmark. In line with Betterment’s investment philosophy, Betterment does not engage in strategies that mirror shorting because of its risky nature and negative expected returns.

Inverse ETFs were therefore eliminated altogether from consideration. The mean the expense ratio of inverse ETFs was 1% per year, while the mean expense ratio of the remaining noninverse ETFs was 0.5%.

Leveraged ETFs

Leveraged ETFs often allow for two or three times the exposure to the underlying indexes using derivatives. Betterment does not hold leveraged ETFs because of their high cost.

The mean expense ratio of the excluded set was 0.96%, while the mean expense ratio of the included set was 0.46%.


The total annual cost of ownership (TACO) is Betterment’s proprietary fund scoring method. TACO takes into account an ETF’s transactional and liquidity costs as well as fund management costs. TACO balances the expense that’s incurred annually (expense ratio) with a transactional cost such as the bid-ask spread.

The weight given to the bid-ask spread, which is incurred only when bought and sold, is calibrated to reflect the average turnover in a Betterment portfolio, which is designed to be bought and held. As such, the bid-ask spread is effectively weighted less than the expense ratio.

In other words, since we do not trade actively in our portfolios, we prefer a fund with an expense ratio of 0.10% and a bid-ask spread of 0.20% to a fund with an expense ratio of 0.20%, and a bid-ask spread of 0.10%.

Within each asset class, we specify a maximum acceptable TACO number. This serves to filter down the total number of funds per asset class to a top-tier set. The maximum TACO threshold differs by asset class. In general, emerging market equity and bond funds have a higher expense than U.S. corporate bond funds. The average expense ratio of the included set was 0.24%, and that of the excluded set was 0.35%.

Average Daily Volume and the 10% ADV Test

We then rank funds based on the liquidity and depth of their market, measured by the average daily volume (ADV) of their trading. ADV measures how deep the market is for a fund, in terms of the amount of shares traded by active buyers and sellers. A fund with a low ADV may have higher-than-expected liquidity costs on days when Betterment needs to buy or sell that fund. We therefore select the top 12 candidates in each component.

In addition to ranking based on ADV, to help ensure Betterment’s orders do not account for more than 10% of the funds’ total daily volume, we create a 10% of ADV test. The test uses an upper bound on expected daily volume in Betterment’s portfolio, and breaks down that volume into the volume per each component. Funds chosen for Betterment’s portfolio cannot exceed the 10% total traded daily volume threshold.

As with previous steps, eliminating thinly traded ETFs reduced expense ratios as well. The excluded set had a mean expense ratio of 0.50%, while the mean expense ratio of the included set was 0.21%.

While this list of our filters and criteria is far from being comprehensive, it gives a flavor for the kinds of characteristics we look for in an ETF.

All in, 30 ETFs met all selection criteria and were placed in the Betterment portfolio.

How We Decide to Change Our Funds

The process above is run quarterly to update our assessments. However, we do not thoughtlessly change our funds immediately when this output changes. By investing customers into a fund, we are potentially committing to them holding it indefinitely, or incurring tax or transaction costs to change it. We must consider if a relative improvement is likely to be permanent, or if a competitor is likely to match the reduction (hooray, competition!).

To decide to make a change, we convene our investment committee and discuss whether we believe such a change is merited based not only on current statistics, but on the trend and behavior of market participants.

We are constantly on the watch for new ETFs that come to market. Meanwhile, we are also monitoring our current portfolio for changes to our ETFs such as increases in expense ratio, tighter bid-ask spreads and shrinking assets under management.

The value of this ETF automation selection process is how easy it is to evaluate the full universe of U.S.-listed ETFs and select the ones that most closely align with our investment philosophy. Through this process, we seek to drive transparency, efficiency and cost-effectiveness on an ongoing basis when re-evaluating our portfolio of ETFs.

This article originally appeared on

About Betterment

Betterment is the largest independent online financial advisor with more than $9 billion in assets under management. The service is designed to help increase customers’ long-term returns and lower taxes for retirement planning, building wealth, and other financial goals. Betterment takes advanced investment strategies and uses technology to deliver them to more than 250,000 customers across its three business lines: direct-to-consumer, Betterment for Advisors, and Betterment for Business. Learn more.

Comments are closed for this post.

Tags: , , , , , , , , ,

About the Author: Dan Egan

Dan Egan is the Director of Behavioral Finance and Investments at Betterment. He has spent his career using behavioral finance to help people make better financial and investment decisions. Dan is a published author of multiple publications related to behavioral economics. He lectures at New York University, London Business School, and the London School of Economics on the topic. Contact Dan at

You can contact Dan via email or follow on Twitter.

3 thoughts on “Eliminating the Black Box: How We Automated ETF Selection

  1. Ellie,

    Great breakdown, and I know we all appreciate the transparency.

    Please point me in the right direction if it exist. How does Betterment determine what allocation to give to each asset class. On those same lines, how does Betterment determine what asset classes to target, i.e. why no growth ETF’s, or market-cap segmentation for Intl ETF’s (are they just too expensive, or is the risk/reward not worth it). And lastly, does Betterment account for ETF’s that share similar underlying assets, i.e. VTI (US Total Market) vs VTV (US Large Value).


    • Hi Lee,

      Our asset allocation decision uses the Black-Litterman model to determine the weightings given to each asset component. The model assumes that asset allocation should be proportional to the market values of the available assets. The market capitalization of each asset proportional to the market cap of all available assets drives the allocation percentages.

      Betterment has a value lean in our ETF selection. We don’t use growth ETFs because we believe that funds that hold stocks deemed to be undervalued in price have an advantage when it comes long term expected returns.

      While we could have chosen market cap segmented ETFs, they entail much higher fees than their returns warrant.

      Lastly, yes Betterment does account for ETFs that share similar underlying assets. However, two markets that share overlapping constituencies can still represent two very different underlying markets. For example, we invest in AGG and LQD so some corporate bonds inevitably go into either ETF. However AGG and LQD differ sufficiently in duration risk and credit risk to both be considered separately in our portfolio.

      Thank you for your questions Lee. I hope that adequately addresses some of your questions.


  2. Thanks for the information. I had always wondered about the selection process, whether there was a lot of data behind it or more of a “I have a gut feeling these stocks will do well” approach.

Comments for this post are closed.

Have something to add or have a question? Email us.