A place for developers to talk about building on Farcaster.

/dev

A month ago, I shared some theoretical work on how vector embeddings could be used to build market predictions. Today, let's dive into some of the integration details behind the analysis, and give a peek at the results. 👀⬇️

NodeJS | TS | Solidity builder 🔵
I like to experiment with FC frames 🔧

🟢 Performance Evaluation

Using a daily CRON job, the system (let’s call it Viktor from now on) starts by evaluating the previous day’s predictions. It fetches the current prices of all selected tokens and records them to track daily performance easily.

🔵 Token Discovery

Next, Viktor uses the Mobula APIs to query all listed tokens (~27,000), then filters them based on:

- Minimum market cap
- Minimum liquidity
- Exclusion of stablecoins
- Compatibility with supported chains (enabling auto-trading via Uniswap’s SwapRouter)

This initial filter leaves us with ~300 tokens on average. These are scored using volatility, volume, liquidity, market cap, and social metrics, then sorted to keep only the most active and liquid ones.

Finally, Viktor performs a check for the presence of ETH or USDC pools on each token’s chain using RPC calls to the Uniswap Factory contract of the token’s chain. After this, the token discovery phase is complete and ready for vector analysis.

🟡 Analysis and cosine similarity search

For each token, Viktor fetches the last 10 days of OHLCV data (Open, High, Low, Close, Volume, which is the underlying data of market charts).
This window is then normalized and converted to natural language and embedded as a vector and compared, via Supabase pgvector, against ~56,000 (at time of writing) training windows stored in memory.

This lets us compute weighted averages for each outcome group, estimate potential returns, and assign a confidence score to the forecast.

Tokens are then ranked based on their potential for positive price movement in the days ahead. 🚀

⚪ Results and conclusion

All daily results are saved on Supabase and can be monitored on an interface (https://viktor-monitor.wakushi.com) so I can track the evolution of performances as I’m improving the system and tweaking some of its rules.
This first version of Viktor was launched after achieving 55-60% of good predictions by comparing past windows analysis against its training vector embedded data on thousand of days.

However it would required multiple month/years of running to make an unbiased conclusion as to this system is reliable, that’s why I’m still working to shorten this delay and increasing its performances. 🔧

In a future post, I’ll share one example of potential improvement leveraging Mobula “smart money” API which tracks the best traders by realized PnL and other useful metrics that Viktor could use to weight some decisions.

Thanks for reading, have a good one 👍
(Happy to hear your thoughts on this :))