Fetch forecasts from prediction markets/forecasting platforms to make them searchable. Integrate these forecasts into other services. https://metaforecast.org/
Go to file
2021-04-10 15:49:20 +02:00
data Made small change to xriskdatabase, and prepared for the creation of a history database 2021-04-10 12:28:32 +02:00
node_modules Refactored application to use a mongodb database 2021-04-08 18:42:48 +02:00
src Fixed bug in addToHistory function 2021-04-10 15:49:20 +02:00
.gitignore Reverting to previous update method 2021-04-09 19:39:03 +02:00
metaforecasts.png Changed about page 2021-03-28 21:13:50 +02:00
package-lock.json Refactored application to use a mongodb database 2021-04-08 18:42:48 +02:00
package.json Changed pointer to index.js 2021-04-08 19:24:31 +02:00
Procfile Changed old password to correct new process variable 2021-04-08 23:29:08 +02:00
README.md Continuing to integrate netlify 2021-04-09 00:17:10 +02:00

What this is

This is a set of libraries and a command line interface that fetches probabilities/forecasts from prediction markets and forecasting platforms.

These forecasts are then used to power a search engine for probabilities, which can be found here (try searching "Trump", "China" or "Semiconductors") (source code here). A json endpoint can be found here.

I also created a search engine using Elicit's IDE, which uses GPT-3 to deliver vastly superior semantic search (as opposed to fuzzy word matching). If you have access to the Elicit IDE, you can use the action "Search Metaforecast database".

How to run

1. Download this repository

git clone https://github.com/QURIresearch/metaforecasts

2. Enter your own process.env variables

The following environment variables are currently needed to run the master branch:

  • MONGODB_URL, a string in the format "mongodb+srv://<username>:<password>@<mongodburl>/?retryWrites=true&w=majority&useNewUrlParser=true&useUnifiedTopology=true"
  • REBUIDNETLIFYHOOKURL, a string in the format "https://api.netlify.com/build_hooks/someprivatestring"
  • CSETFORETELL_COOKIE
  • GOODJUDGMENTOPENCOOKIE
  • HYPERMINDCOOKIE

The cookie formats can be found in src/input/privatekeys_example.json; these session cookies are necessary to query CSET-foretell, Good Judgment Open and Hypermind. You can get these cookies by creating an account in said platforms and then making and inspecting a request (e.g., by making a prediction, or browsing questions). After doing this, you should create the environment variables.

Alternatively, for fewer complications, have a look at the commandlineinterface branch, which instead of requiring environment variables only requires a src/privatekeys.json, in the same format as its src/privatekeys_example.json. Its disadvantages are that the command line tool in the commandlineinterface is more difficult to integrate with other services.

3. Actually run

From the top level directory, enter: npm run start

What are "stars" and how are they computed

Star ratings—e.g. ★★★☆☆—are an indicator of the quality of an aggregate forecast for a question. These ratings currently try to reflect my own best judgment and the best judgment of forecasting experts I've asked, based on our collective experience forecasting on these platforms. Thus, stars have a strong subjective component which could be formalized and refined in the future. You can see the code used to decide how many stars to assign here

With regards the quality, I am most uncertain about Smarkets, Hypermind, Ladbrokes and WilliamHill, as I haven't used them as much. Also note that, whatever other redeeming features they might have, prediction markets rarely go above 95% or below 5%.

Various notes

  • Right now, I'm fetching only a couple of common properties, such as the title, url, platform, whether a question is binary (yes/no), its percentage, and the number of forecasts. However, the code contains more fields commented out, such as trade volume, liquidity, etc.
  • A note as to quality: Tentatively, Good Judgment >> Good Judgment Open ~ Metaculus > CSET > PredictIt ~> Polymarket >> Elicit > Omen.
  • I'm not really sure where Hypermind falls in that spectrum.
  • For elicit and metaculus, this library currently filters questions with <10 predictions.
  • Omen does have very few active predictions at the moment; this is not a mistake.