1. Heard on the Street – Transaction Cost or Liquidity Cost
A portfolio manager of a major hedge fund once revealed that their annual execution cost was estimated beyond $100 MM. Assuming the gross annual trading is $100 Billion, it implies an average cost of 10 bps (basis points), or a cost of $10 for trading every $10,000.
Communication like this goes on every day in the financial world, especially within cost-conscious institutions such as hedge funds, retirement funds, wealth or asset management firms, and major broker-dealers. This cost item has been tagged with different names: execution cost, market impact cost, transaction cost, or liquidity cost. They may differ in different contexts but always share the same core component – the inevitable and often cumulative cost due to information leakage and liquidity taking when executing large trades in public markets.
We shall converge to the usage of “transaction cost” and “transaction cost models” (TCM) in this discussion. In particular, we do not consider logistic costs like exchange fees, clearing fees, or broker commission fees. These are defined by rulebooks and directly observable, and do not require modeling.
2. An Academic Thought Experiment – Arising of Market Impact & Transaction Cost
In complex markets, how a particular trade generates market impact can be quite obscure. It is helpful to examine how market impact and related costs could arise in a much simplified market, in which a single trading participant (P1) faces off with a single market maker (MM).
- Assume that P1 wants to purchase 30 lots of stock XYZYX, or 3,000 shares before the market closes at 4:00 pm. Now it is 3:30 pm.
- Assume that at 3:30 pm, the MM quotes at $99(Bid) | $101(Ask), both with 4 lots.
- P1 may first take out the 4 lots at $101(A). This initial fish bite prompts the MM to adjust higher the B/A at $100(B) | $102(A) from $99(B) | $101(A), but with only 2 lots at both.
- P1 waits for 3 minutes, before taking out the 2 lots at $102(A). Now the MM has a firmer belief that more fish bites are to come, and hence aggressively adjusts the B/A at $101(B) | $106(A), with 1 lot at both.
- “Come on!” P1 sighs, followed by a long pause for 7 minutes. The MM starts to doubt her own judgment – “Oh, my overreaction – no more fish bites.” The MM hence adjusts the B/A lower and narrower at $101(B) | $103(A), with 4 lots at both.
- P1 Aha’s and takes out the 4 lots at $103(A) with little hesitation. The MM now quotes at $103(A)|$106(B)…
- P1 waits, but it is 3:42 pm now, 3:43, 3:44, … With each minute elapsing in the rear mirror, time pressure starts to build up before the market closes at 4:00 pm. Hence P1 becomes more aggressive in taking out…
- (Market Impact) P1’s purchasing has caused the perceived market to jump from $99(B)|$101(A) to $109(B)|$111(A). This is the market impact of the trade. Using the mid prices, the market impact can be quantified by (110-100)/100=10.0%.
- (Transaction Cost) Compared with the perceived market (mid) price of $100 at the arrival (3:30 pm), the actual per share price has been $105. The difference of $5 is the shortfall or transaction cost. In percentage, it is (105-100)/100=5.0%.
3. Universal Challenge of Low Signal-to-Noise Ratio (SNR) for TCM Modeling
A real market (e.g., US Equities) usually consists of multiple market makers and numerous retail or institutional market participants. Many institutional players ( e.g., quant funds) may actively apply intraday trading strategies for opportunistic liquidity provision or taking. Passive investors (e.g., retirement funds) could also generate substantial volume or price waves due to fund rebalancing or portfolio repositioning. Intricately intertwined, together they create a rolling market dynamics too complex to decipher.
Unlike the academic thought experiment in the preceding section, it is virtually impossible to segregate the net market effects of a moderately participated trade in such complex markets. Take the stock of Apple Inc. (AAPL) for instance. Yahoo Finance shows that the average daily volume (ADV) is about $12.0 Billion. Thus a trade of $1.2 MM, say, is only at 0.01% participation of volume (PoV), and its direct market impact is barely noticeable. This is because on average one merely executes a single order after the market completes 10,000, assuming individual orders are more or less of the same notional size. Herein we always assume that such a trade is executed professionally via a scheduling algorithm like VWAP or TWAP. A naive single-shot market order could cause a flash crash. A big trade has to be executed via much smaller “sliced” orders.
Thus for liquid stocks or securities, most moderate trades result in very low signal-to-noise ratios (SNR) for transaction costs. Like a tender cherry flower petal swallowed by a torrent, the floating path has little to do with her own properties but more with the surrounding currents.
At the other end of the liquidity spectrum, the transaction costs of illiquid names are known to be highly idiosyncratic and hard to forecast. Market impacts may depend sensitively on many factors, including whether a trade is done in the lit, dark, both markets, or via OTC block trading. In extreme cases, trading may actually degenerate towards the idealized thought experiment in the preceding section, i.e., a single trader vs. a single designated market maker (DMM). For illiquid names, the DMM’s quoting algorithm may sensitively depend on her limited inventory position. Therefore, the transaction costs to purchase 3000 shares of an illiquid stock whose average daily volume is only 5000 shares, say, may vary substantially from one trading instance to another.
In either scenarios, the huge background variations or noises always overshadow the target signal, i.e., the market impact of a particular trade.
These low SNRs become the main obstacles for any professionals working on TCM modeling. It does not brighten up even when a major investment bank or broker-dealer possesses hundreds of millions of proprietary sample trades. The R-squared scores of their TCMs are invariably gloomy, with some as low as the single-digit level.
Even Machine Learning has not been able to turn the tide, whether using gradient-boosting trees or deep neural networks. Training errors only slightly improve, but generalization power on test samples always deteriorates.
4. DQT’s Innovative Approach and Model: DQT-TCM-US-EQ
DQT faces off against the curse of Low SNRs in an innovative way. The model is based on our non-public and non-published proprietary research results. Such efforts reflect the spirit of “Deep” in “Deep QuanTech,” even for areas traditionally deemed well understood or well done in the financial world. As revealed earlier, the perception of “well done” is no more than just a wishful illusion. TCM research is far from maturity.
(Asset Class) For this initial phase, DQT’s TCM models only cover equities, which is the most liquid and well-regulated asset class, as compared with credits or commodities for example.
(Pilot Release) The pilot release is the Model DQT-TCM-US-EQ, which covers about 3000 top US stocks and 600+ top US ETFs. These US equity securities are often the most wanted by the investment world.
(Scoping) Due to the high idiosyncrasies associated with trading illiquid securities, we have narrowed down our focus only to relatively more liquid securities. Our model only requires a security to have a more transparent price-discovery process. Hence in the future, we shall be able to release TCM models for many liquid asset classes and securities, like rates, futures, liquid FICC ETFs, or liquid cryptos for example.
(Unique Approach) Our TCM models are based on our non-published/non-public proprietary research and framework. It excels not merely via Machine Learning (ML), as one would naively bet on in this AI age. It is not uncommon that anything tagged with the fashion name of ML or AI can be re-marketed as a brand new product. But TCM is different. Machine Learning does not improve much as revealed in the preceding section.
We bravely deviate from the conventional TCM modeling approach, which relies on millions of individual VWAP/TWAP (etc) trades that are proprietarily owned by leading investment banks, buy-side execution desks, or agency trading houses. Our approach is a bold attempt to break the curse of very low SNRs, and we believe it has worked. DQT’s TCM models only rely on publicly available market data, e.g., the TAQ (Trade and Quote) data and the aggregated intraday data, etc. Whether in academia or on Wall Street, it has never been thought possible!
(Try Out the Model GUI) Please try out the GUI for Model DQT-TCM-US-EQ.
We wrap up with Einstein’s famous quote – “Everything should be made as simple as possible, but not simpler.” Yes, a simple and clean GUI designed just for you on the front end, with all necessary outputs that cannot be made simpler, and with a deeply innovative framework and model on the back end. This is our TCM offering.
“Go Deep” – has always been the very spirit of Deep QuanTech.
5. Disclaimer
- (Model Risk) As for any model-based estimations, please be aware of the model risk (as in FED SR11-7).
- (Model Usage) GUI results are calculated for your reference, not to advise on specific investments. Under extra business review or coupled with other quant tools, they could be used to (a) assess the level of brokerage fees or markup offerings, (b) create pre-trade analytics and execution expectations, (c) assess liquidation costs under normal or stressed market environments, (d) create portfolios or select individual stocks that optimally balance alphas, betas, risks, and costs, and (e) contribute to many other pre-trade analytical tasks.
- (Benchmarking) GUI results should be better compared with those produced by your other vendors (like the Bloomberg Terminal) or brokerage services (like from your brokers). One should always dive into the scenarios when such benchmarking gaps become concerning, including exercising more business judgment.
- (Mean and Variance) The current TCM model outputs the mean transaction cost C for a given trade. In reality for each real trade, the actual cost structure could be C + N, where N is the noise or fluctuating cost due to the innate Brownian motion of the market (i.e., the market risk). In a T-trending market, it could be also C+N+T.
- (Third Parties) In this white paper, any mentions of third-party names, e.g. Apple Inc., do not represent an outbound business endorsement from DQT on them, nor an inbound business endorsement on DQT from these parties. Instead, they are cited as examples because they are household names on Main or Wall Street.