Don’t Look Back At ETF Trading Costs

August 12, 2021

Laszlo HolloLaszlo Hollo is a vice president in MSCI’s Risk Management Research team based in Budapest. He leads the team focusing on liquidity-related regulations and transaction cost modeling.

His recent research identifies transaction costs as a leading indicator of the relationship between ETF prices and net asset value (NAV), and offers insights on how advisors can better manage the risk involved with discrepancies between these two figures.

The following transcript has been edited for clarity and brevity. What are transaction costs? How do they show up within an ETF trade?

Laszlo Hollo: I referred to transaction cost as the difference between the average trade price and the preliminary fair price. Let’s consider exchange-traded products, but similar argumentation applies to OTC [over the counter] markets too.

If you’re selling a small equity holding, say 100 shares, you may execute the trade at or close to the prevailing bid price. If we mark the equity at the midprice, which is the average of the best bid and ask quote, then your trade price is lower than the midprice by about one-half of the bid/ask spread. The transaction cost of this trade is half bid/ask.

If you’re trading a larger size, you may not execute the full trade at the best price, but your order will be matched with limit orders deeper in the order book.

As such, if you’re selling, your average trade price will be even lower than the best bid. Your transaction cost became larger than the half bid/ask spread. So the transaction costs depend on the size you’re trading and also on the liquidity of the market. Taking only the bid/ask spread into account may underestimate the transaction costs.

In the analysis presented in my blog post. I used MSCI’s liquidity risk model, LiquidityMetrics, to estimate transaction cost for trading the basket of constituent stocks. What factors go into MSCI’s estimate of transaction costs?

Hollo: In our LiquidityMetrics framework, there are multiple [points of] market data we’re using in our liquidity models.

What we’re using in the calibration methodology is exchange data—what were the volumes, what was the price, what was the bid/offer spread during the last couple of weeks or months. And Virtu [the data vendor] also uses data that it calls the peer data universe that is connected to buy-side institutions.

Whenever the buy-side institution submits a trade order, they can identify this trade order from the exchange data. They can calibrate what’s called “implementation shortfall.” Implementation shortfall is basically the difference between your execution price and the trade price. This is basically the transaction cost.

That all goes into the calibration. What comes out [is] the liquidity model. It tells you what the expected transaction costs are for a given size. If you’re trading 100 shares, it’s going to be close to the path of the bid/ask spread, but if you’re transacting more [than 100 shares], it will tell how steep your transaction costs will increase if you increase the trade size. Why are transaction costs a better way to estimate price/NAV risk compared to using the historical differences between the two?

Hollo: Historical price/NAV difference is a backward-looking measure. If you’d like to calibrate the risk of the price/NAV difference, then by risk, you’re usually referring to an upper limit. If you’d like to use the historical approach, one approach could be selecting a high percentile of this price/NAV difference for a standard period.

If you’re using the 99th percentile, you need to have a long enough observation window to reliably calibrate this historical percentile. That makes these historical observations slow to react. If there’s a market dislocation, you’re still using data from the past.

Compared to this approach that we were looking at, it’s not using the historical price/NAV difference. What we use from historical data is [an estimate of] how large of flows we might expect. That’s our only assumption in this research piece. We estimated these flows using the 99th percentile of observed fund flows.

We then use the model to estimate the transaction costs for selling 5% of this fund on the equity market. Our liquidity model is more reactive than the history of observed price/NAV differences, and that can make our estimation more reactive to changing market conditions. The research you did primarily used thematic ETFs. Would the transaction cost model work for other categories, like fixed income?

Hollo: In principle, the framework would work for other ETFs. For equity ETFs, though, the problem is simpler. For equity ETFs, the redemption basket is a vertical slice of the ETF. However, for bond ETFs, the redemption basket or the creation basket can be quite different from the vertical slice of the ETFs.

The difference between the creation and redemption basket compared to the cross section of the ETF may make this more complex. But in principle, the same approach would give you some estimate on how large the price/NAV difference may get. Are there market environments where the transaction cost model would be less reliable?

Hollo: When there's a crisis period, there are usually other factors that contribute to the price/NAV difference. It’s not only the transaction costs; there are a handful of other factors. The creation and redemption mechanism is a fairly complex mechanism that authorized participants implement on their side.

First, the timing of the exchange: If you have an ETF that’s listed in the U.S. but is highly invested in Japanese equities, the equity price cannot really move. Still, there are transactions on the ETF market, so there might be discrepancies between the [price and the] underlying NAV that are frozen.

Another factor that may contribute, especially during a period of crisis, is inefficiency of price discovery. The pricing of the underlying securities may become less obvious. The ETF is used as a proxy for a price discovery, because the ETF price usually reacts faster and provides more reliable pricing than the underlying security. That can also contribute to the price/NAV difference. If you can’t value the NAV with high certainty, but you can value the ETF price with lower uncertainty, the difference will still have some uncertainty. What's the main takeaway advisors should get from your research? And how can they use it to optimize their trading practices?

Hollo: The main insight is that they can get a more reactive view of the risk between the price of the ETF and the NAV than by looking at historical differences. If a market participant is interested in modeling this risk, relying on some sort of liquidity model might give them a fairly good estimate of this. As a first approach, they could look at bid/ask spreads or trading volumes on a typical ETF.

This data is usually quite accessible. Still, looking at only the bid/ask spread or the trading volumes won't give you a full picture. As I mentioned earlier, if you’re transacting larger sizes, you can’t really expect the trade to be close to the bid or the ask. The bid/ask spread only gives you a lower limit on the transaction costs.

Looking at volume is a widely used approach to estimate the underlying equity markets liquidity. However, during the COVID crisis, we saw that trading volumes went up.

But liquidity didn’t improve during this time, because at the same time, transaction costs also went up: Bid/ask spreads were widening and the additional market impact of medium-size trades further increased. This means that the costs for selling the same size increased during the pandemic outbreak.

So looking at volumes may not give you a full picture of how easy would it be for you to trade on the underlying market. That’s the reason we were looking at our integrated model that looks at volumes and cost at the same time.

Our model gave fairly sensible results for the selected ETFs in terms of comparing our transaction cost estimates and the observed price/NAV differences.

Jessica Ferringer can be reached at [email protected]

Find your next ETF

Reset All