Forecast Accuracy & Track Record MCP: forecasts vs actuals, prediction markets, archived reports
Check a forecast against what actually happened, as MCP tools: institutional forecast track records, economic releases with forecast vs actual, earnings estimates vs results, prediction-market prices and leaderboards, plus archived web pages and news for what was said at the time. One hosted endpoint, one Actor per job.
Hosted remote MCP server (Streamable HTTP) on Apify's MCP gateway. Client config plus a small open-source stdio proxy server. No Actor source.
Disclosure: entries marked "(maintained by us)" are Actors I maintain on Apify Store (publisher publicrecords). Every other Actor is independent, made by its own publisher and picked on merit (Store usage, ratings and recent maintenance). This repo is not affiliated with those publishers or with the sites the Actors read.
MCP endpoint
Auth: sign in with Apify (OAuth) when your client prompts, or send an Apify API token as a Bearer token. Runs are billed on your own Apify account under each Actor's own Store pricing; see each Store page.
Tools
The Apify gateway also exposes helper tools to read run results (get-actor-run, get-dataset-items). Want fewer tools? Drop any Actor from the tools= list.
Who it's for
- Analysts and journalists: how accurate has this institution's forecast been before?
- Macro and markets: did the release beat or miss consensus?
- Researchers: what did prediction markets price before the outcome?
- Fact-checkers: what did a page or the news say on a given date?
Example prompts
- The IMF projects
<country> GDP growth of <x>% next year. How accurate have its past projections been?
- Show this week's economic releases for the US with forecast vs actual.
- How often has
<ticker> beaten its EPS estimate over the last 8 quarters?
- What did
<url> look like in the Wayback Machine around <date>?
Cursor / Claude Code / any remote-MCP client
Claude Code:
Claude Desktop (local bridge)
Local stdio server (open source)
This repo also ships a small MCP server in src/ for clients that only run local stdio servers. It answers initialize and tools/list locally from tools.json, which is generated from each Actor's public input schema and checked for drift in CI and weekly, so any client or directory can inspect the tools with no token. tools/call is forwarded to the hosted Apify MCP endpoint above with your own APIFY_TOKEN; without a token a call returns a clear error instead of running anything. The token is only sent in the Authorization header and is never logged.
Or from a clone:
Optional env: APIFY_ACTORS (comma-separated subset of the Actors above) and APIFY_MCP_URL (full endpoint override).
Tests run against an in-memory mock server, with no network and no token: npm test. Requires Node.js 18+.
Python (uvx/pip)
A matching Python version of this server lives in python/, built on the official mcp Python SDK: same tools, same APIFY_TOKEN, same hosted endpoint. Requires Python 3.10+.
Or with pip: pip install forecast-accuracy-mcp, then run forecast-accuracy-mcp. Python tests: cd python && pip install -e ".[test]" && env -u APIFY_TOKEN python -m pytest.
See also
Registry
Official MCP Registry name: io.github.GeminiGeorge22/forecast-accuracy-mcp
License
MIT, for this documentation, config and the proxy server code in src/. Each Actor belongs to its publisher.
