Bid-Ask Spreads Error Bars and Trusting Your Own Research
A market-maker’s bid-ask spread can be understood as (not only) the cost of immediacy but as an operationalization of epistemic distance between far is the present’s partial knowledge and the future’s eventual consensus. That is to say the bid-ask spread reflects, not merely overhead or transaction fees but rather exists as a dynamic confidence interval over the collective belief state of regarding value, risk, information asymmetry, and higher order uncertainty.
Secondly, I present an alternative model in which a simplifying true value exists, toward which the market settles by the end of trading.
I aim to address common questions such as:
- What are liquidity providers actually doing, and why should they make money?
- What is the role of informed versus uninformed participants?
- Who bears the cost of payment for order flow?
- Why does speed matter?
- Who bears the cost of insider trading if, after all, it brings more information to the market?
In the case of prediction markets, bid-ask spreads are seemingly a closer analogy to statistical error bars. There are, of course, caveats: while typical error bars indicate a 95% probability that the true value lies within a given range (meaning one would not be surprised by any value in that range), the interpretation of the bid-ask spread is more along the lines of "I would not prima facie trade against anyone within this range." This subtle difference accounts for the fact that the generation process for market orders differs from generic random variables which would miss the key fact that it typically entails a decision step wherein a participant believes a transaction will make them better off, which could reflect a change in the variable's true value or exogenous factors, like routine retirement contributions. Unless one is paying for clean order flow or are otherwise informed about your counterparty's information informedness, you are forced to assume the worst: namely, that your counterparty possesses information you lack. However since the market consists of both informed and uninformed order flow, losses to informed traders can be offset by the gains from providing competitive prices to uninformed traders. Returning to the core question, what is the actual interpretation of the bid-ask spread? - It depends on who your counterparties are and what you assume about the order generation process that accounts for their behavior. - It depends on how confident you are regarding the true value of a variable.
Suppose you are a market maker and the only participant with access to the order book or information about it (e.g., observing which orders get filled). Further assume that no new information enters the market, in the sense that while participants possess private information, no new exogenous events affect the asset's volume, that is information entering the market is only mediated by orderflow.
Every participant independently makes and evaluation of what the “true value” is and then decides how much capital to commit based on factors like total assets, confidence, or portfolio balance.
Consider a market model with 100 participants, each with a fixed order size of 1 unit (50 buyers and 50 sellers). The market maker wants to execute any order which yields a profit net of transaction costs (assumed here to be zero), and may choose to hold off on execution until better information arrives. Because they have exclusive access to the order book, they do not fear being front-run, but they must still account for the risk that a participant's information is superior to their own.
If compatible orders are not automatically executed without your intervention, you can simply match bids higher than asks and collect guaranteed profits.However, if compatible orders execute automatically, you must earn your return by taking risk under strict operational constraints.
To formalize this model, we define the following constraints: - Inventory constraint: Inventory $I_t$ must satisfy $|I_t| \leq I_{\max}$. - Fill probability: $P_{\text{fill}}(p)$ denotes the probability that a limit order quoted at price $p$ is executed. - Terminal inventory: $I_T = 0$, requiring all positions to be closed by the end of the trading horizon.
Under these conditions, filling an incoming buy order at price $p_{\text{buy}}$ has positive expected return if
$$ p_{\text{buy}} > \mathbb{E}\left[\min_{t>t_0}p_{\text{ask},t}\mid\text{fill}\right]. $$
The expected value of the best available buy price after an additional observation is
$$ \text{Expected Value} = \mathbb{E}\left[ \max\left( \text{current\_best\_buy}, \text{newly\_sampled\_buy} \right) \right] $$
But what does this represent in intuitive terms?
Assuming a "True Value" signifying the market-clearing price once all orders are placed, then if you knew the clearing price in advance, your bid-ask spread would center precisely on that “True Value”. However the situation is complicated by the requirement to quote a bid-ask spread continuously. Consider the specific case with 100 total orders, where we evaluate the spread at the 99th order for the 100th participant. Assume there is no outstanding net inventory and that all the prior 98 orders have arrived, and that the final incoming order is a buy order. We would fill our best sell offer if the expected value of the buy price exceeds that offer price. Because only one sample remains, the value is $\mathbb{E}[\max(\text{current\_best\_buy}, \text{newly\_sampled\_buy})]$. The distribution of this maximum is a mixture distribution consisting of a Dirac delta at the current best buy (corresponding to the newly sample buy being lower) and the tail of a normal distribution (corresponding to the new sample being higher).
How does this relate to the “True Price”? If we fill an ask, the settlement price matches this distribution. Otherwise, the final settlement price will be the best ask if the buy order exceeds it, or the price of the last settlement if it does not.
Does this approach maximize the execution price for the buy order? If a buy order is filled at step 100, this strategy maximizes the expected value of that settlement price.
What is our best bid? It is the expected value of the best bid available in the immediate future. We compress time to bring our best estimate of future prices into the present. Conversely, our best ask is the expected value of the best ask that will become available. When trading against informed counterparties, this expectation is conditioned on the fact that they have placed an order. Consider the 98th order. There are three potential combinations for the remaining two orders, which simplify without loss of generality to (buy, sell) and (buy, buy). In the (buy, buy) case, we fill any sell order below the expected maximum of those two incoming buy values. In case of (buy, sell), filling an ask follows the same logic as the single-buy case, with the exception that an unseen ask may arrive first with a 50% probability. If that ask is better than the current best ask, filling the order early incurs an opportunity cost equal to the price difference.
We can return now to the question of "What is the bid–ask spread really encoding?"
When market participants are confident that new orders will not surprise them, (i.e. when information is symmetric), when the order flow is uninformed, when volatility is low; then the expected value of future best bids and asks collapses toward the current price, and the spread tightens.
When on the other hand uncertainty rises, e.g. because some orders are informed (or are believed to be), volatility increases, or the distribution of possible future prices widens, then the expected value gap between best-bid and best-ask widens accordingly The spread reflects this epistemic gap in addition to structural inventory and immediacy costs.
To examine this in another manner, suppose we set up a game where a room full of people are designated as bidders and choosers. Participants guess the longitude of a city and receive payouts proportional to how many degrees of longitude west the true location is relative to their guess (losing money if they err east), provided another participant takes the opposite side. Periodically, participants are randomly selected to challenge a fact, representing liquidity entering or exiting the market. This models liquidity provision and taking on the bid side. While "true price" is an abstract concept because we cannot rollout all possible future states and interest rates to calculate the empirical expected value, we can simplify by assuming definite facts correspond to true prices—in this analogy, the longitudes of cities.
From this perspective, the presence of market participants buying order flow increases uncertainty, as liquidity providers must state stronger beliefs conditioned on an informed participant trading against them. You would offer much wider ranges in a room full of geography experts than in a room of general participants.
If we refine the game so bidders are paid to err west and askers to err east, we see a concrete instantiation of spreads. Amateurs quoting standing offers fear that overbidding too far east will cause informed participants to trade against them, resulting in a loss, whereas underbidding may result in no trade without incurring a direct loss.
Regarding insider trading, suppose participants enter the room with maps to make money. In the short term, they adjust prices toward the true city locations, enabling liquidity takers to transact more effectively. However, as other participants realize they are trading against informed counterparties, they will withdraw (or be more conservative in) their bids in fear of adverse selection. If it gets out of hand market participants may stop quoting all together, leaving the market bounds indeterminate. The risk of informed traders entering the market reduces the willingness of liquidity providers to quote tight spreads, in proportion to their belief of the likelihood and severity of having them as counterparties.
A liquidity provider underwrites a belief for the duration an order remains live. If new information arrives–a narrowing spread on Bologna informing participants about the plausible longitude of Rome—failing to update an outstanding bid on Rome represents a failure to incorporate new evidence/information. Likewise if you have not been able to update on all currently known information your uncertainty will of necessity be higher. So speed matters on both sides. You want to know all the information that is there and you don’t want to have an order that gets executed because adverse information comes to light.
Hidden liquidity on the other hand takes a weaker epistemic stance because its execution is conditional on a liquidity taker filling the order without prior knowledge of its presence. Leaving counterparties less informed regarding total available market depth.
Participation also shapes the market, casual/uninformed participants may accept an expected loss for entertainment, whereas professional geographers are presumed to need expected returns to participate. But one group’s gain is the other’s loss and casual participants will leave (or not show up) if losses become excessive.
But the spread is more than just a disagreement about price it is a readoff from the belief state of the participants on all of these factors and of their beliefs about the market itself e.g. how informed their counterparties are how fast information is changing, how much information they are contributing (or will be believed to be contributing) by posting the liquidity that they are posting, about the relative informedness not just of the average market participant but of the weighted average of who is likely to be their counterparty.
The bid-ask spread is an expression of the market’s higher order uncertainty: it is not only the value of an asset that must be accounted for but who and under what circumstances will take the other side, what do they know, what do they know about what you know, and on to ever high-order theory-of-mind.
So back to the questions we opened with, what are liquidity providers doing? They are underwriting beliefs for the duration their liquidity is on the market providing bounds on the collective belief state. What roles do informed and uninformed participants take? Uninformed participants incentivise research while participating for other reasons, informed participants contribute information to improve the collective belief state (Grossman and Stiglitz (1980) and Black (1986)). Who bears the costs of payment for order flow? Market participants who need to accept wider belief bounds/bid ask spreads on the open market as a result (Easley, Kiefer, & O'Hara, 1996). Does speed matter? Yes, since it allows market makers to update quotes before informed counterparties exploit stale beliefs. Who bears the costs of insider trading, it is contributing information to the market after all? While insider trades can locally move prices towards being more informed expectation of insider trading mean lower market efficiency and in the limit arbitrarily wide spreads due to the expectation of arbitrarily informed counterparties.The spread is not merely a transaction tax, but a dynamic equilibrium compensating liquidity providers for inventory risk and information asymmetry, in the forms of inventory risk (Stoll, 1978; Ho & Stoll, 1981), immediacy costs (Demsetz, 1968) and adverse selection (Bagehot, 1971).
Thanks to my father Robert Litzenberger for editing and guidance.
References
- Bagehot, W. (1971). The Only Game in Town. Financial Analysts Journal, 27(2), 12–14.
- Black, F. (1986). Noise. The Journal of Finance, 41(3), 528–543.
- Demsetz, H. (1968). The Cost of Transacting. The Quarterly Journal of Economics, 82(1), 33–53.
- Easley, D., Kiefer, N. M., & O'Hara, M. (1996). Cream-Skimming or Profit-Sharing? The Spread of Order Flow. The Journal of Finance, 51(3), 811–833.
- Grossman, S. J., & Stiglitz, J. E. (1980). On the Impossibility of Informationally Efficient Markets. The American Economic Review, 70(3), 393–408.
- Ho, T., & Stoll, H. R. (1981). Optimal dealer pricing under transactions and return uncertainty. Journal of Financial Economics, 9(1), 47–73.
- Stoll, H. R. (1978). The Supply of Dealer Services in Securities Markets. The Journal of Finance, 33(4), 1133–1151.
-
read model ↩