Monte Carlo Simulations for Predictions
dyokomizo opened this issue · 0 comments
A tool to build Bayesian networks out of statistical variables to build scenarios and simulate their outcomes using Monte Carlo.
Scenario analysis is the main focus of this tool, e.g. comparing two different scenarios in multiple ways.
Guesstimate is similar. It supports integration using SLURP (from the SIPmath Standard). It should be possible to prototype using it.
How can we make optimal decisions when outcomes aren’t deterministic? Traditional optimization techniques can’t apply directly because they don’t address uncertainty. Instead, we need a framework for specifying the random behaviors of a system, and algorithms that can handle randomness.
Markov decision processes offer such a framework and come with a suite of optimization algorithms. Haskell is a great learning tool for making mathematical abstractions more concrete, so this talk will teach Markov decision processes by designing a Haskell library for working with them.
Tikhon Jelvis – Reasoning under Uncertainty
Ought provides a library to access Metaculus through its API: Ergo.
They also developed Elicit that turns predictions into continuous distributions.
A preliminary roadmap:
- monte carlo simulation for a single variable with known distribution
- multiple independent variables
- dependent variables
- save on dropbox/google drive
- scenarios analysis (e.g. grouping results, visualizations)
- import from metaculus.com
- import from goodjudgement.com
- autosave