Introduction

Optimism is an L2 Protocol for Ethereum that has been doing quite a lot of exciting experiments to guide the future of its protocol. One of the first pieces was the announcement of the airdrop and who would qualify for it. While noble in intent, some have claimed that the airdrop didn’t go as well as intended. As of 10/13, 88% of airdrop wallets decreased their $OP position by 99%. To stay optimistic about Optimism is that there is a second airdrop happening and their ecosystem is thriving. What if there was a way to learn from past on-chain data, to predict future trends in the airdrop with some on-chain sleuthing and data science? In this analysis, we look at what went well, what went wrong, and how users interacted with $OP Optimism before and after the token's launch.

Methodology

Based on previous analysis on the $HOP token, I took the existing code/model I had developed it and translated it to the needs of Optimism Data was provided by Flipside and integrated with a Jupyter Notebook via ShroomDK code for the OP analysis can be found here:

To further enhance/cater the model to the needs of Optimism I added the following:

• Labels to showcase OP airdrop qualifiers: (DAO Voter, User Priced Out of Ethereum, Gitcoin Donor, Multisig Signer, Repeated Optimism User, Bridge Optimism User). This was devised using the airdrop amount a wallet was given (too lazy to do the query! for each component)

• Optimism Activity: # of Txs, Days since the first tx, days since last tx # of interactions to/from tx ratio, etc. Cut off was the announcement of OP airdrop on April 26th

• Ethereum Activity: # of Txs, Days since the first tx, days since last tx # of interactions to/from tx ratio, etc. . Cut off was the announcement of OP airdrop on April 26th

• Optimism Bridging Activity: How many bridges using the OP bridge, volume, average size, etc. Cut off was the announcement of OP airdrop on April 26th

This was then analyzed and processed through a classification model to determine if an address was likely to dump the airdrop

Insights

• Only ~2100 wallets (1% of airdropped wallets) have increased their $OP position and .

Breaking down the 7,30 to present-day windows of claiming airdrops we can see that a majority of wallets dumped $OP within the first week of the airdrop.

Breaking down the OP qualifications there is also an interesting pattern. The highest users to dump were those priced out of Ethereum. Perhaps this is due to them not being interested in Optimism in the first place. Those who has used Optimism repeatedly were the lowest, which makes sense. “Optimism Users” as of current writing has the second highest then followed by gitcoin donors. My thought around this is that these groups was where Sybils/Farmers were most likely to lurk as its the easiest to do

Surprisingly, 70% of wallets at one point and time delegated their $OP. Is due to them trying to receive further rewards or they are actively interested in the network?

Once again multi qualifiers and repeated optimism users were most likely to delegate. Gitcoin donors , Multisig signers, and Optimism users had the three lowest.

In terms of those that dumped previous airdrops, some of the most popular tokens to dump were ENS and X2Y2 (~10% of users qualifying). Those that dumped a previous airdrop were more likely to dump $OP.

Classification Model Insights

The model resulted in a ~67% Area under the curve for predicting if the wallet would dump. This is lowered than the $Hop model prediction earlier but it shows that dumpers can still be identified. Additionally, when the model was over-sampled (decreased the number of dumpers to a similar size) The model’s area under the curve was 74%

Like the $Hop Model, total txs played an important part in model prediction (high amount of txs were Farmers and traders, low may be bots) Activity Span on OP, total fees on OP, distinct interactions on Optimism also plan important factors. This further highlights the case where if users were ACTUALLY using Optimism then they were less likely to dump (unlike DAO voters, Gitcoin donors, potiential Bridge farmers). Dumping past airdrops also had a higher likelihood of dumping $OP.

While the model results are hard to determine for many users if they would/wouldn’t dump based on chain activity, I wanted to root out those addresses that have a high conviction of dumping so that further protocols would be able to take these results and identify unwanted behaviors or Sybil. By looking at those wallets that had an 80% or greater chance of dumping $OP as well as those that used the bridge method as this was the most likely chance of being a farmer / Sybil before the announcement of the airdrop. About 16% of the dataset has a high conviction of dumping in the dataset and this result in an accuracy of ~86% Over 2000 wallets were identified through this method

One of the interesting things about this was that by identifying Sybil certain behaviors emerged:

1. Sybil would receive crypto from some random address, that would send crypto to 50+ wallets. Surprisingly many of these wallets had ens addresses

2. They would send money to Optimism, Arbitrum, and zksync bridge. This is important as these wallets can be rooted out of airdrops for Arbtirum and zksync

3. Once they were on Optimism, they did nothing

4)They would transfer any remaining crypto out of their wallets to another wallet. Rinse and repeat

It’s hard to penalize those who are on the fence and actually using the product. But in the case of farmers and sybils, I think the model does a good job of rooting out bad behavior. Manual verification may be needed still to double-check addresses.

Post OP Activity

While dumping $OP is one thing, did wallets actually use Optimism after? Some claim that airdrops are a marketing stunt and help reward early users of potential risks. If this is the case, then shouldn’t airdrop wallets still be using the product afterward? I used the same query for the Optimism activity and looked at if dumpers vs non-dumpers were using the network afterward (Query was all activity after 4/26). From the data, it can be seen that dumpers had lower distinct active days and weeks compared to non-dumpers. For total transactions, dumpers vs non-dumpers had the same amount of transactions. However, the top 25% of non-dumpers were more active in transactions than dumpers. The optimistic piece is that some dumpers were still interacting with Optimism, showing a sign of success for the network

When looking at the span of Optimism activity (day since the claim vs the last transaction on the network). 46% of dumped wallets were inactive after 30 days after claiming with 35% of wallets being inactive after just 3 days of claiming OP. This number is quite small for those that didn’t dump with 9% being inactive after just 3 days and 10% being inactive after 30. This is a very small difference.

Breaking down the $OP qualification categories, users that had not used Optimism were more likely to not use the network after. 69% of users priced out of Ethereum were inactive with that number increasing to 78%. Multi-sig signers also had the highest inactivity on Optimism. Is this because they weren’t interested in the network from the start?

Those that had repeatedly used Optimism had the lowest network turnover (as well as “Optimism UIsers” and those who qualified through the airdrop from multiple means). This drives a point that those who are repeatedly using the network should be awarded for consistent use.

Conclusion

Optimism is a vibrant network and has a lot going for it. It is constantly experimenting with new methods. From this analysis, the following can be gathered about the airdrop:

1. Majority of users dumped the token within using it, however those that didn’t were loyal to the network

2. Those that hadn’t used the network before (Users price out of Ethereum, Dao Voters, Gitcoin Donors, multisig signers) were more likely to dump the token and not use the network after

3. Those that had used Optimsim especially repeatedly were less likely to dump and still continue to contribute to the Optimism network

4. Machine learning classification can be used to detect dumpers. While for some users it may be tough to tell, it was shown to have high accuracy for where the model predicted a high likely hood of dumping. This was also translated over to sybil/farming detection

5. While a lot of users dumped $OP, the majority of users did delegate their $OP at some point