Alpha Architect: Internal Costs Of ETFs

September 26, 2018

[This ETF Industry Perspective is sponsored by Alpha Architect.]

Alpha Architect, as a firm, embraces the concept of academically driven investment practices; in particular, factors such as value and momentum. The issuer’s website is a trove of research and analysis. Here, CIO Jack Vogel expands upon the topic of “internal costs”—trading costs, in particular—of ETFs, a subject ETF.com Managing Director Dave Nadig wrote about in a 2014 blog. Investors may find that the costs associated with an ETF are more complicated than they previously assumed.

ETF.com: I wanted to dig into the trading costs associated with running a factor-based strategy. How do trading costs differ across ETFs?
Jack: I would say there are two ways in which the trading costs will differ.

First is rebalance frequency. Most passive ETFs (think S&P 500) will only rebalance once a year, whenever the index updates. When one turns toward more active ETFs, specifically those that rebalance more frequently (such as quarterly), the trading costs will be higher, simply because they rebalance more often.

Second, one should examine the weighting of the positions within an ETF. Passive options (again, think S&P 500) generally market-capitalization-weight the positions, which gives higher weights to larger stocks, which (generally) happen to be more liquid. The more liquid the stocks, the lower the trading costs are to buy/sell these stocks. Alternative strategies may equal-weight or overweight positions based on some ranking (such as value or momentum), which can cause higher weights to less liquid stocks, thereby increasing the trading costs.

ETF.com: It sounds like passive ETFs generally have the lowest trading costs, while more active strategies have higher trading costs, is that accurate?
Jack: Yes, that is accurate, in general.

ETF.com: Digging into more active or “smart beta”-type strategies, is it possible that these trading costs will eliminate any premium associated with the strategy, such as the value or momentum premium?
Jack: To give some background on the topic, the “value” or “momentum” premiums that fund companies reference were originally tested by academics assuming zero transaction costs. So, a natural question arises—do these anomalies exist after accounting for transaction costs?

Academics have been studying this for some time now.

The microstructure literature attempts to estimate trading costs using “TAQ” (trade and quote) data, which has detailed information such as bid/ask spreads, volume and trades for each transaction down to the nanosecond—it really is a neat, and large, data set.

From here, academics have attempted to calculate the true costs of implementing all the standard factor portfolios, such as value, momentum, quality, etc. The goal of their studies is to identify how much money one can invest in a factor strategy before the trading costs outweigh the benefits, such as the value or momentum premiums. Clearly this is important for investors to know!

However, estimating true trading costs involves creating a model. And depending on the model one uses, the estimates of how much money can be invested in a strategy will vary. For example, academic models, based on papers here, here and here, argue that trading costs would eliminate the long/short momentum premium, around $5 billion within the U.S., which is not that large.

Alternatively, practitioners (and academics, hereafter called “pracademics”) from BlackRock and AQR have studies (found here and here) using their own trading data, which have the momentum premium existing on a much larger scale, around $50 billion to $300 billion.

ETF.com: How can the estimates be so different?
Jack: As I mentioned above, to estimate trading costs, one needs to build a model. Most of the purely academic studies assume a linear trading-cost model, whereas the pracademic (meaning they combine academic and practitioner perspectives) studies assume more complex, but real-world, trading models. For example, the pracademic models assume smarter trading methods. As such, they come to massively different conclusions.

ETF.com: Who do you think is right in this fight?
Jack: I think the pracademic papers have better estimates. The main reason can be found in the updated AQR paper, titled “Trading Costs” (our take is here) and the evidence in Table IX in that paper. This table examines the trading costs of trading standard indices, such as the S&P 500 and Russell 2000.

From the live ETF data (Vanguard S&P 500 Admiral Fund and the iShares Russell 2000 ETF), the authors back out the true trading costs to be 4.72 and 12.87 basis points (bps) respectively. Using the pracademic model, the paper estimates the trading costs to run the funds to be 4.81 and 12.36 bps. However, using a standard linear TAQ model (as used in most of the academic papers), the linear models predict trading costs to be 29.79 and 155.79 bps.

As you can see, the linear models used in the academic papers can be off by magnitudes of seven to 10 times. Since the pracademic model produces numbers more in line with the true trading costs (4.27 vs. 4.81 and 12.87 vs. 12.36), I am more inclined to believe their models.

ETF.com: Very interesting to see the differences in the data, simply by changing the model.
Jack: I agree; small differences in assumptions cause major differences.

In fact, an interesting idea is to attempt to estimate trading costs by “inferring” the trading costs, as opposed to modeling them using the TAQ data. Doing so, in effect, eliminates assumptions on how to model the data. Papers on this topic can be found here and here. High level, these studies claim that, after transaction costs, most smart-beta strategies yield little to no factor premiums; in effect, saying trading costs destroy the factor premiums.

However, like the issue with models, changing small assumptions has a massive effect on the results. As I point out here (about halfway through), if one assumes that funds (1) closet-index, or (2) change factors, one can eliminate the premiums to factors portfolios, without assuming any transaction costs. Thus, while I like the new approach, I would disagree with the conclusions from the papers on “inferred” trading costs.

ETF.com: What about the capacity constraints? I find that interesting given there are hundreds of smart-beta ETFs.
Jack: As mentioned above, there are capacity constraints to different factor-based strategies, as the BlackRock and AQR papers highlight.

Let’s say, for argument’s sake, that there is $50 billion invested in momentum funds (the capacity outlined in the AQR paper)—is that premium gone? Possibly, but we always need to place any strategy in context—an interesting paper that does this is by David Blitz, found here. This paper finds that, in aggregate, all ETFs basically rebuild the market, while having no tilt toward any factor!

Technically speaking, all ETFs in aggregate statistically significantly load on the market factor (t-stat = 64.31) and have an insignificant loading on all the standard factors (size, value, momentum and volatility). So, given all the smart-beta ETFs available, investors (through their ETF selections) have literally just re-created the market-cap-weighted portfolio.

ETF.com: So, big picture, what should investors know about factor investing and trading costs?
Jack: Overall, there are a few takeaways. First is that trading costs, by definition, will reduce returns. Second, while there are limits to how much one can invest in a smart-beta strategy, the capacity should be put into context. Third, investors interested in learning more should dig into the details of the above studies and can read my summary on factor investing and trading costs.

Find your next ETF

CLEAR FILTER