Monitoring FUD with AI: Which DeFi Telegram Group Better Reflects Negative Sentiments?

2022 was littered with negative events in crypto, and many of them were intensively discussed over social media channels before the actual event happened. Monitoring potential rekt events is imperative for risk analysis of lending markets, and in the RiskDAO we constantly work on automating the monitoring process using natural language processing AI techniques.

DeFi has its own internal slang and jargon, and therefore training a sentiment AI model is not an easy task. In this blog post we describe the experiments we did around UST depeging and FTX meltdown.

Meet LobsterDAO and the REKT telegram channels. The first group is dedicated to research around economics, governance, tech and DeFi projects in general. The second group is the dark web of DeFi journalism, and continuously discuss past and future hacks and meltdowns.

We decided to analyze the sentiment of both groups during two important situations in 2022 — the Terra collapse and the FTX bankruptcy — to see which one is better at detecting FUD (fear, uncertainty, and doubt). Keep reading to find out what we discovered.

Methodology

The aim of this analysis is to compare sentiment between the LobsterDAO and Rekt channels.

In order to maintain consistency in our analysis, we make a key differentiation between the two channels.

• For Rekt, which is primarily focused on exploits, hacks, and vulnerabilities, we assume a negativity bias in conversations. Therefore, we counted only mentions of FTX/Luna without filtering for sentiment data.
• For LobsterDAO, on the other hand, we utilize Natural Language Processing (NLP) to determine the sentiment of messages that contain relevant keywords.

Upon charting the results, we observe that the scores move inversely: In the case of the Rekt channel, a rising trend line indicates deteriorating sentiment (number of mentions goes up), while a declining trend line in the LobsterDAO channel suggests the same (sentiment score goes down).

FTX bankruptcy

The defining moment for the FTX bankruptcy was the abrupt halting of withdrawals. This also coincided with announcements of a potential bailout by Binance (which fell through).

The analysis of the FTX saga reveals that the LobsterDAO sentiment graph declined much earlier than the Rekt one went up. LobsterDAO sentiment had already begun deteriorating before FTX halted withdrawals, while negative sentiment in the Rekt channel only gained momentum immediately prior to the “news” of FTX halting withdrawals.

In the graph below, the green line describes the average sentiment of messages that contained FTX related keywords on LobsterDAO. The blue line describes the number of mentions of FTX related keywords in the REKT group. The vertical bar stands for the time of the last on-chain withdrawal transaction from FTX.

Luna/UST

In the case of $LUNA and $UST, the project’s fate was sealed when the tokens entered a downward spiral from which they could not recover.

We apply the same analytical methodology from our prior investigation to the Luna/UST implosion: Our analysis indicates that LobsterDAO proved to be a more adept indicator of the severity of the situation as compared to Rekt, with the green chart revealing an earlier drop-off compared to the blue chart which only started to increase as the event had unfolded.

Conclusion

The above analysis suggests that FUD discussions begin to show at mainstream forums such as LobsterDAO and it takes them time until they trickle to FUD dedicated forums such as REKT. On the negative side, LobsterDAO group has much more content, and thus is hard to follow. This is where the automated tools that we develop at RiskDAO become handy.

This analysis should be considered as an experiment that although not statistically meaningful, holds some weight about the predictive power of communities. Future iterations of this analysis can include rigorous scientific approaches and larger sample sizes.