Deepmint
Market calls, recorded before the fact and scored against real prices after it.
What is Deepmint?
Deepmint keeps score of stock-market predictions.
When someone makes a call such as "Nvidia will be higher in 3 months", Deepmint stores it as a claim: which stock, which direction, for how long, and the stock's most recent closing price. Deepmint never edits the claim afterwards. When the 3 months are up, it looks up the real closing price, works out whether the call was right and how much it would have gained or lost, and adds that result to the author's track record.
Over many claims this builds a scorecard: how often someone is right, how much their calls return, and where they rank against everyone else. Deepmint also combines everyone's recent calls on a stock into a weighted consensus.
Today Deepmint covers seven large US technology stocks, the "Magnificent 7": Apple, Microsoft, Alphabet, Amazon, Nvidia, Meta and Tesla. That is a deliberate choice: more stocks will follow, but only once the core product has matured. What's live today has the details.
Where to find it: the app is live, ahead of its public launch, at www.deepmint.ai. Sign up there, then log a call with New Claim, in the top bar or on the dashboard. Ticker pages, Learn and claim pages are public; everything else needs an account.
Who is it for?
| Who | What they can do today |
|---|
| Guides: analysts and research firms | Their calls are collected for them. Deepmint reads their published posts, or an archive of past posts, and an AI model extracts the calls. A Guide needs no account. An analyst can also apply from Settings, where the landing page's "Apply as a Guide" leads: an admin approves them as a Guide, or makes the Guide profile Deepmint already built for them their account. |
| Players: everyone who signs up | Log your own claims in the web app. They are checked against real prices by the same rules as Guides' claims. You appear on the leaderboard next to the Guides. You can also run practice ("paper") portfolios with virtual money. |
| Followers | This is not a separate account type. Any Player can follow Guides and other Players, read their claims in a "Following" feed, keep a watchlist of stocks, and use "Mirror Signals" to copy someone's calls into a paper portfolio. A mirror copies a call when it goes live, if the call is dated within the last 48 hours: an old rating or a late approval is not copied, and neither is a backfill or a neutral call. |
| Admins | Admin rights are granted by the maintainer in Clerk. Admins approve or reject claims held for review and Guide applications, switch Guides' feeds on, manage the list of stocks, issue API keys and delete accounts. |
| Developers and AI agents | Read scores, consensus and leaderboards through a REST API, using a key issued by an admin. AI agents can use the same data through the MCP server at /api/mcp, with the same kind of key. |
Why does it exist?
Market predictions are everywhere: TV, newsletters, social media, bank research notes. Almost nobody keeps score. Calls are vague about timing, wrong ones are quietly forgotten, and right ones get repeated. You cannot tell skill from luck or from selective memory.
Deepmint's answer:
- Record it before the outcome. A claim is stored with its date before anyone knows how it turns out, and its scored move starts at the first market close after it was made. The exceptions are backfilled history (past posts loaded from an archive, dated to when they were published) and analyst ratings, which are dated to the day the rating was issued.
- Never rewrite it. Claims are append-only. The only change allowed is an admin approving or rejecting a claim that was held for review. Database triggers enforce this, for events, notes and outcomes too.
- Put a deadline on every call. Every claim has a horizon between 1 day and 1 year, so "eventually" does not count.
- Use real prices and plain statistics. Results come from end-of-day closing prices (Polygon), and scores are ordinary statistics such as hit rate, average return and the Sharpe ratio. No AI is involved in scoring. The only AI in Deepmint is the model that reads text to pull out claims.
How does it work?
flowchart TD
P["Player logs a claim in the web app"] --> A
G["Guide posts collected from feeds or an archive"] --> L["AI model extracts claims and checks the evidence"]
L -->|"confident, horizon quoted"| A["Active claim"]
L -->|"not sure"| R["Admin review queue"]
R -->|"approve"| A
R -->|"reject"| X["Rejected, kept on record"]
A -->|"horizon ends"| M["Markout: compare with the closing price"]
M --> S["Scoring: hit rate, returns, EIV"]
S --> B["Leaderboard"]
S --> C["Consensus for each stock"]
This is the beginning of the README. Read the rest at the original. Read the full README ↗
