The Daily Finance Brief
Every weekday before the market opens, I get an email with the finance Twitter ideas worth a closer look, ranked by whose past calls held up.
Why I built it
Twitter is full of people with stock picks. The harder question is: who should you actually listen to?
I started tracking a group of accounts I follow that regularly make stock calls, measuring how those calls held up over time. Now, instead of scrolling through my feed or dozens of accounts every morning, I get one briefing with the accounts, and the calls, worth paying attention to.
Sample briefing
High Conviction Ideas
Sector Themes
Watchlist
How It Works
I started with 15 accounts whose ideas I've traded on. The script looks at who they retweet, quote, or mention favorably to find other accounts that keep coming up, and it now tracks 45+.
Discover
Find new accounts through the seed accounts' connections
15 → 45+ accountsCollect
Pull only tweets posted since the last run
~90% fewer API callsExtract
A cheap first pass drops tweets that aren't stock calls. A second pulls out the ticker, long or short, and how confident the author sounds.
6,080 tweets → 1,956 callsScore
Compare each call's return to the S&P 500 after 30, 60 and 90 days
An account's score firms up as its record grows (Bayesian)Deliver
Send the email: top ideas, sector themes, watchlist
Sent daily via Resend before the openNetwork Discovery
Decisions That Shaped the Project
Local model → gpt-4o-mini
I started with a local model through Ollama because it was free. At 5–15 seconds a tweet, 6,000 tweets took most of a day. gpt-4o-mini did the same job in about 50 minutes for $0.60.
Two-Pass Extraction
Most tweets aren't stock calls, so a cheap yes/no check runs first. Only the ~30% that pass get the full extraction, which cut cost by about 70%.
What I use it for
I use the email to decide which companies to read about that morning. I don't trade off it directly. A 30–90 day window is too short to prove an edge, and tweets are too messy by themselves to attempt an automated strategy.
Does the Scoring Work?
If the scores mean anything, the top-ranked accounts should do clearly better than the bottom ones. In the first run, they did:
The briefing ranks each idea by its author's record, so calls from accounts like the top one come first.
Open Questions
Seed Bias
Every tracked account traces back to 15 I picked, so the list is at least somewhat representative of my picks.
Out-of-Sample Validation
The scores were built and tested on the same period. The real test is out-of-sample: rank the accounts on one stretch of time, then freeze the scores and test whether they hold up in a later period.
Risk Adjustment
A call that dropped 30% before ending up 10% counts the same as one that rose steadily. Drawdown isn't measured yet.