Blockin.ai: LP Tracking, LP Clustering Analysis

Author: Blockin.ai

I. Statistical Sample

The statistical sample consists of all 85,195 LPs from the establishment of the Uniswap V3 protocol until October 27, 2022. Among them, 1,066 LPs transferred their positions to another address after their initial injection and did not continue to hold those positions, so they were excluded, leaving a remaining sample of 84,129.

This article derives hourly LP characteristic indicators based on all position-related events and hourly aggregated revenue metrics for LPs and their positions from on-chain metadata, used for LP characteristic distribution statistics, establishing LP clustering, and user profiling.

II. LP Characteristic Distribution Statistics

(1) Distribution of LP's Historical Maximum Number of Pools and Positions
The maximum number of pools held simultaneously by LPs, the maximum number of positions, and the maximum position in a single pool were counted. The distribution is shown in Figure 1-1, revealing that over 66,000 LPs choose to provide liquidity in only one pool at the same time, over 66,000 LPs hold only one position in a single pool at the same time, and over 57,000 LPs hold only one position in Uni-V3 at the same time, indicating that most LPs are single-pool single-position holders, with only a small number of LPs choosing to diversify their investments.

LPs are divided into single-position holders and multi-position holders (singleposition / multipositions). An example of the position distribution of multi-position holders within a single pool is shown in Figure 1-2:

LP Address: 0x741aa7cfb2c7bf2a1e7d4da2e3df6a56ca4131f3;

POOL Address: 0x8ad599c3a0ff1de082011efddc58f1908eb6e6d8;

Figure 2-1 Distribution of LP's Historical Maximum Number of Pools, Maximum Number of Positions, and Maximum Position in a Single Pool

Figure 2-2 Example of Position Distribution of Multi-Position Holders in a Single Pool

(2) Statistical Analysis of Total Mining Duration and Rebalancing Frequency of LPs

The total mining days (total_days) refer to the time from when an LP first acquired a position to the most recent complete withdrawal of that position. The distribution of total mining days is mainly concentrated between 0-100 days and follows a long-tail distribution; it was found that there are about 20,000 LPs with total mining days exceeding 200 days, and about 16,000 LPs with total mining days less than or equal to 3 days.

The number of rebalancing actions (operate_cnt) refers to the number of mint/burn actions performed by LPs on their positions. For statistical convenience, all rebalancing actions of a single position are counted at the last owner of the position. As shown in Figure 1-3, the number of rebalancing actions is mainly concentrated at 2 and 3, with about 54,000 LPs having rebalancing actions less than or equal to 3.

Since a single withdrawn position must have at least two rebalancing actions (one injection and one withdrawal), LPs with operatecnt <= 2 are considered single-operation LPs, LPs with 2 < operatecnt <= 10 are considered multi-operation LPs, and LPs with operate_cnt > 10 are considered many-operation LPs. The distribution of rebalancing action categories is shown in Figure 1-4.

Figure 2-3 Density Distribution of LP's Total Mining Duration and Number of Rebalancing Actions

Figure 2-4 Distribution of Rebalancing Action Categories

(3) Distribution of Last Withdrawal Time and Yield of LPs

From the distribution of the last withdrawal month of withdrawn LPs, the highest number of LPs completely withdrew in October 2022, while the number of LPs completely withdrawing in other months is relatively even, with the least number of complete withdrawals occurring in February 2022.

Figure 2-4 Distribution of LPs by Last Withdrawal Month

(4) Distribution of LP Position Ranges

By calculating the average width of the position ranges and the upper and lower bounds of LPs in terms of token1, we observe the tendency of LPs to set position range widths. The standard width is calculated as: (upper limit - lower limit) / current price.

The average range width of LPs shows a bimodal distribution, with the first peak around 0.15, indicating that some LPs believe the pool's exchange price will fluctuate within 15%; the second peak around 1.5 indicates that some LPs tend to have a wider distribution, primarily concentrated between 0-2.

LPs that consider the upper range width to be more than twice the lower range have a positive price tendency, while those that consider the lower range width to be more than twice the upper range have a negative price tendency; others are neutral. As shown, LPs setting ranges are mostly neutral, with those having a positive tendency slightly fewer than neutral, and those with a negative tendency being relatively few.

Figure 2-5 Distribution of LPs by Position Range Tendency

III. Feature Selection

(1) Multicollinearity and Correlation Analysis of Features
Through multicollinearity analysis, the VIF values of features were calculated, and features with a maximum VIF value greater than 10 were sequentially removed. The removed features include: 'hismaxnetvalueualld', 'avgnetvalueu', 'avgposage', 'posage_75percent'.

The correlation coefficient matrix of numerical variables was calculated, and a heatmap of the correlation coefficients was created. Correlation coefficients above 0.7 are considered highly correlated features, and these will not be selected in subsequent variable filtering.

Figure 3-1 Correlation Coefficient Heatmap

(2) Important Feature Binning

The historical total annualized return of LPs was divided into a binary variable [1, 0], where a positive return is assigned a value of 1 and a negative return a value of 0. The final result was obtained through equal-frequency binning, filtering features with an IV value of no less than 0.2 and high correlation features, resulting in the following important features:

From the feature binning results, LPs with larger net positions, later last withdrawal times, shorter average rebalancing times, and more rebalancing actions have a lower proportion of negative returns within the group; the distribution of net annualized return in the feature binning also supports this conclusion.

Table 3-1 Feature Binning IV

Table 3-2 Feature Binning Results

Figure 3-2 Box Plot of LP Net Return Distribution Based on Feature Binning

Based on the goal of selecting LPs for following, it is necessary to choose LPs that have recently made operations, so the most recent withdrawal time of LPs is not used as a feature indicator.

The proportion of LPs with positive returns was calculated for the remaining three groups of features, and a heatmap was drawn. It was found that the feature distinction of the number of rebalancing actions is not obvious, while LPs with a maximum net position > 3631.76 & average rebalancing time < 76 show significantly better overall returns, which also aligns with the characteristics of smart money. Therefore, subsequent LP tracking will select from LPs that meet these conditions.

There are a total of 32,117 LPs that meet the above conditions. To find active addresses, we also require that the most recent withdrawal time be within the last 15 days, cumulative mining days exceed 30 days, and days in the pool exceed 20 days, resulting in 2,201 LPs.

Figure 3-3 Heatmap of Positive Return Distribution of Important Feature Binning LPs

(3) LP Operational Style Feature Binning

Figure 3-3 Box Plot of LPs Classified by Operational Style