Optimal Bidding Strategy for Maker Auctions [research paper]

Executive Summary
This paper proposes a bidding strategy based on “participation costs." Using this strategy, we calculate the maximum recommended bidding price for a sample of auctions. 75% of winning bids in the sample were unprofitable, as the discount from market price did not cover the participation costs. This behavior may be caused by inexperienced participants, or by altruistic actors who desire to keep the auction system running smoothly. An example calculation may be found here.

Theory
What price should keepers bid at in Maker auctions? A new working paper, entitled “Optimal Bidding Strategy for Maker Auctions” attempts to answer this question. We are posting this paper to allow for comments, prior to uploading to ArXiv.

Maker liquidations auctions are primarily common-value auctions: everyone has the same valuation for the auction collateral, which is the market price of the collateral. However, there are two private-value components:

  1. Alternative usage of collateral
  2. Costs to participate in the auction

The paper focuses primarily on “participation costs,” which are defined as the following:

  1. Transaction fees: Gas to dent/deal, join/exit from the Vat, and trade on Uniswap
  2. Conversion costs: Slippage when converting WETH to DAI
  3. Cost of capital: Implicit required return (discount rate) on capital held for bidding.

Other costs are incurred as a keeper (hardware, maintenance, etc.), but these are not incurred on chain, and are therefore not included in the analysis.

The paper proposes that participation costs can be optimized to their lowest possible point, by adjusting the minimum/maximum capital balance held for auction bidding. Once participation costs are optimized, bidders can determine the minimum discount from market price that they need to still make a profit. Alternatively stated, bidders can determine the maximum price they should bid at for each auction.

A total of 155 auctions, from March 23 through July 28, 2020, were analyzed. The analysis first optimized the participation costs to the lowest possible total, based on the auction parameters (tab) and the network conditions (gas price, and current ETH/DAI price). The analysis then compared the optimal bidding price to the actual bidding price that won each auction.

Findings
The analysis found that 75% of auctions were won at prices that did not cover participation costs. That is, the actual price was above the optimal price, which meant that the discount from the current market price was too small for the winning bidders to gain a profit, according our model of participation costs.
Latex table

The gap in optimal vs. actual price is most prevalent in bids of $1,000 or less (83% of these auctions were won at a loss). The two charts below show the optimal discount vs. the actual discount / markup, relative to the market price at the time of the auction.

scatter0to1000
scatter1000up

Several theories may be proposed to explain the differences observed.

  1. Indifference to cost of capital: This cost is implicit, and some participants may not be thinking about their required rate of return (estimated at 40%).
  2. Inexperienced actors: Some participants may not be accounting for the entirety of gas / slippage costs.
  3. Altruistic actors: Some participants may bid at a loss in the auctions, but they still benefit from having the Maker Protocol run smoothly (e.g. no repeat of Black Thursday). Even though they are losing money on paper, their total utility from the system is positive.

The paper (see here) covers the theory and findings in greater detail.

What does this mean for Liquidations 2.0?
The strategy proposed in this paper will still apply for Liquidations 2.0, but the definition of participation costs will certainly change under the new system. For example, one proposal under the new system is to lower participation costs by adding the ability to “flash-mint.” The full effects of the new system will need to be modeled after the system design has been finalized.

Bonus feature
A Google spreadsheet has been created to show the model used to calculate participation costs. The sheet is included for any keepers that want to test their own strategy, or try to improve on the calculations.

Other notes

  • I am the first listed author in this paper.
  • The research team does not hold any MKR or DAI. The team has also never participated in a mainnet auction (only on the Kovan network).
6 Likes

This is fascinating. Great work.

1 Like

liquidation information is something I am interested in generally. I have not tried to pull new data but after the ttl and lot size changes and during the bull market in ETH (other assets had runs as well) it is interesting to look at these things.

So I have questions regarding the above analysis.

Do you have a breakdown of number of bidders, and numbers of bids. My liquidation data suggested there was a strong correlation of number of bidders to auction performance. There was also an element of data suggesting human access and errors were proving to be in favor of vault holders (i.e. significant overbids). So knowing how much over theoretical collateral return there was generally is important.

Costs to bid (win or lose) are a bid deal - especially with gas costs. This was one thing i was very interested in generally. How much total cost in gas for each auction was needed just to complete an auction relative to the value of the auction (i.e. as a percentage) and then compare this to amount of collateral return. it is my expectation that the gas used by all auction participants is now becoming significant relative to the collateral returned

It is my hope the liquidation system 2.0 will mitigate a lot of the gas costs to participate.

I am also interested generally whether some keepers are purposely outbidding others as lix suggested creating a kind of exhaustion effect on other keepers trying to bid. Ao how often keepers won vs. lost and whether this had any effect on keepers continuing to participate or not (i.e. keeper losing exhaustion and give up - ala keeper fatigue). Could we isolate keepers that might be making the higher bids to exhaust others in the hopes they drop out of the liquidation bidding process… (i.e. has this happened and is it successful).

I am also interested in a break down of the above information based on collateral facility because it was very clear that BAT auctions were doing worse than ETH ones generall (had less participation) so I wondered if this was also true of other collateral types. (Are there distinct difference in results based on the collateral types - sampling may be thin on some of these).

155 auctions in total since Mar 23. Not a lot of data there and a very significant rise in gas prices. It would be interesting to see some of the above data contrasted against the average gas price during the auction. This was the key reason I was interested in accumulating total costs for all participants in the auctions and comparing this to the size of auction and winning price as well as collateral returned.

You are exactly correctly, according to the perspectives of this model. As gas prices rise, the percentage discount should also rise, because it costs more to bid. These effects should be seen most significantly in small bids, because the fixed gas costs represent a high percentage of the total bid value.

The effect of the version 2 auctions would be a fascinating topic for further research.

This is an interesting idea, as an auction corollary of predatory pricing. That is, Keeper A could price bids so high that Keeper B is discouraged from participating.

While I could conjecture this occurring, I would also consider the counterpoint: once the prices fall back into profitable territory, there are few entry barriers to deter Keeper B from jumping back into bidding. As a result, Keeper A would never make a profit, and would likely lose money in this strategy. Considering this fact ex ante, it would be irrational for Keeper A to engage in this strategy.

Also, we ran into the following issue with address-level analyses (per appendix of the paper): “A bidder could easily use multiple addresses, which means that any analysis of bidding history by address would be unable to capture, with certainty, the full behavior of individual bidders.”

The sampling is indeed quite thin for non-ETH liquidations. The aggregate value of non-ETH liquidations (BAT, ZRX, WBTC, etc.) was approximately $18,000 during our sample period, as compared to $1.3M in ETH liquidations during the same time period.

In order to capture the most significant events, we focused on the ETH liquidation auctions only.

We were intentionally narrow in our research focus, as we wanted to document the full equations for this bidding strategy, and then compare this result to historical data. In the context of the formal paper, we didn’t need to add many “dashboard” type analytics.

Having said that, we collected a significant amount of data behind the scenes. For example, we have a SQL database of all auctions events from inception (November 2019) up until the Liquidation 1.1 update (July 28, 2020). From that data, we have a leaderboard that shows the aggregate results for every bidder - totals for lot won, tab paid, gas fees paid, and gross and net profit. See below for a snippet.

The full dataset wasn’t needed in the final research paper. However, if this type of data is useful to the project overall, then we are happy to share relevant information. With all the data in place, it is relatively trivial to calculate the total gas fees for all participants. Other participation costs (e.g. trading fees/slippage on Uniswap) would need to be estimated separately, as we did in the paper.

1 Like