Balancer V2 Pools Trading Fee Methodology
As Fernando mentioned in his recent blog post, we’ve partnered with Balancer Labs to optimize fees for Balancer V2 pools. Our platform allows us to continuously re-run our optimization model and change fees over time to improve liquidity provider (LP) returns. While we intend to iterate on our model as market conditions change, we thought it would be helpful to share how we came to our initial fee recommendations.
Pool Selection Criteria
In selecting pools for the migration to V2, our objective was to migrate as much liquidity as possible with some tradeoff towards active usage. From the set of all Balancer pools, we considered the candidate set of two token finalized pools with:
- 1M+ USD in liquidity
- $100K+ 30D volume, $20K+ 5D volume
- At least 3 liquidity providers
- Trading fees between 5 and 100 basis points (bps)
For each pair of tokens represented by our remaining candidates, if there were multiple pools present, we considered the most active pool by total liquidity and volume, irrespective of weights. We made exceptions to this rule for the 50/50 BAL/WETH and DAI/WETH pools, despite their more active 80/20 counterparts, due to their importance and the gradual easing of rewards from V1 to V2. For similar reasons we also included the 50/50 BAL/DAI pool.
Due to their prominence in DeFi the WETH pools with UNI, LINK, USDC, WBTC, DAI, SNX, ENJ, COMP, MKR, BAT, ZRX, YFI, UMA and BAL were included by default.
From the remaining candidates we selected the top 15 pools according to the following metric
where L is the pool liquidity, Vn represents the volume over the past n days, F represents the total fees collected, and MC represents the market cap of the base token. We chose this metric to allow for each factor to contribute substantially while also reducing the impact of large outliers in any particular factor.
This left us with the following list of 32 pools, with their associated V1 fees:
Model Context and Assumptions
Our fee recommendations are aimed towards dynamically maximizing LP returns throughout changes in market conditions. Balancer rewards represent a substantial share of liquidity provider incentives so an auxiliary focus is to help Balancer achieve better KPIs such as total value locked (TVL) and volume which lead to BAL appreciation, boosting LP returns and ultimately leading to a healthier Balancer ecosystem.
Our analysis showed that LPs are reasonably sophisticated: top tier BAL, WBTC, and other major DeFi pools all converged to similar aggregate APYs (rewards + fees + impermanent loss (IL)) within their respective sets of pools with the same token pair despite these pools differing in weights and fees.
Thus, in modeling future pool liquidity we assumed that external liquidity would move in to bring V2 reward APY back towards the reward equilibrium seen presently in V1 pools. We also assume that V1 pools that receive no rewards in V2 will eventually migrate liquidity into corresponding V2 pools. Reducing the number of overlapping pools works well with the gas savings of Balancer V2 where multi-hop trades are cheaper with deeper more coalesced pools leading to a state with less fee cannibalization, volume fragmentation and fewer LP coordination games.
In considering the effect of fee adjustments on Balancer, lower fees may lead to increased trading interest in the protocol but our analysis suggests that the majority of trades on Balancer are arbitrage volume. As such we direct our primary focus towards the welfare of LPs but will adjust accordingly as the volume profile changes.
As a final note, we formed our best estimates for how rewards would change from V1 to V2 based on community feedback in the forums and from discussions with the Balancer team. This and other inputs of our model will be continuously refreshed to stay up to date with any changes. Having more on chain data in the coming weeks will also help us better iterate on our model and associated assumptions.
Simulation Model
In order to better understand how fees affect LP returns, we conducted a large scale agent based simulation of trading behavior on Balancer.
For each pair of tokens, we consider a range of possible trading fees under a variety of market conditions.
For each pool and each fee under consideration, we simulate hundreds of matrices of price trajectories fit to historical market data, using hourly time steps over the course of a month.
In pools without stablecoins, we use a multivariate geometric Brownian motion process to produce correlated price paths. When simulating stablecoin prices, we use an Ornstein-Uhlenbeck mean reversion process to simulate price behavior on external markets.
We then use Gauntlet’s SDK to simulate realistic behavior for a variety of actors engaging with the Balancer protocol.
At each timestep we simulate the behavior of arbitrageurs, who compare external pricing information to spot prices on Balancer in order to find opportunities that would allow them to profit by swapping coins in the simulated pool. When such opportunities present themselves, arbitrage agents find the trade size that will maximize their profit subject to a number of conditions regarding expected slippage, gas prices, and external fees.
We also simulate natural trading behavior based on previously observed volume data for trades seen in the pools under consideration.
After thousands of simulations, we then choose fees for each pool which maximize a metric of LP returns (fees + IL) with losses penalized quadratically.
Fee Model Factors
In addition to running the simulations as described above, we also used a number of heuristic models to guide fee selection.
LP Breakeven ROI
We estimate the breakeven fee for LP returns, under the assumptions that:
- Future liquidity is affected by future rewards
- Migrated capital from other Balancer pools
- External capital from other DEXs
- Circulating token supply not previously used for LPing
2. Annualized IL remains roughly consistent with that observed in historical data.
3. Trading volume scales as a function of liquidity
- Capped by size of DEX market
The fee is such that:
future volume * fee + future rewards = future IL
where if rewards exceed IL this breakeven fee defaults to 0.
Volatility
We take the realized 30D volatility of each pool, normalizing to map to fee reductions for pools with low volatility and fee increases for those with higher volatility. This is in line with the latest changes in Uniswap V3 with their 3 tiered fee system corresponding to different levels of expected pair volatility.
Impermanent Loss
We use the realized 30D impermanent loss of each pool and the expected impermanent loss incurred from our simulations to produce an estimate of future impermanent loss. This is then normalized to map to fee reductions for pools with low IL and increases for those with high IL.
Organic Volume Share
Through a combination of simulation and historical data, we estimate the percentage of trade volume that can be attributed to either normal traders or arbitrageurs. In general, pools that see most of their volume coming from arbitrage trades are more receptive to higher fees than retail traders, at the expense of pools more loosely tracking external prices.
Pool Utilization
Normalizing for market positions, when the pool utilization on Balancer is higher than Uniswap and Sushiswap a fee increase is suggested, otherwise a fee decrease. This corresponds to bringing the pool closer to the overall DEX supply demand equilibrium.
Circulating Supply Share
Our analysis shows that liquid pools with a larger share of the overall token circulating supply have a lower fee APY. This effect is relatively independent of market cap and possibly due to a couple factors. One is that APYs spikes can be more extreme when supply is too scarce or restricted to meet demand. Another is that DEX demand may lag increases in supply from external sources not previously on DEXs coming in to mine rewards. This may be relevant in liquidity rushing into smaller pools with V2 rewards increases depressing fee APY before volume picks up.
Rewards
Pools that have their rewards increased from V1 to V2 skew towards a fee reduction with the added compensation displacing a possible drop in fee income. Pools that have their rewards reduced correspondingly are assigned a fee increase.
Change Magnitude
Large fee changes should be made in collaboration with the existing Balancer community, and thus we use the prior V1 fee as a baseline and shape our additive updates to be smaller and gradual as opposed to sharp and sudden.
For each fee model listed above, we calculate a recommended fee change. In order to limit the size of model fee updates, we use a hyperbolic tangent to trim updates with an absolute value larger than our maximum fee update, and use a linear combination of these to produce the overall fee recommendation.
Technical Specification of Fee Model
We will update the weights and factors in our model accordingly as we get more feedback and data from Balancer V2 in the coming weeks.
—
We hope this provides context on how pool fees are set. We’re looking forward to the launch of Balancer V2 in the coming weeks.
(Special thanks to Nick Borg, Nick Cannon and Tarun for their help on this post)
