Philotica
Probability forecasts with statistical models and AI Delphi
Start new thread

Philotica - Probability forecasts with statistical models and AI Delphi

by•
Philotica Forecast Studio combines statistical models for US gas storage and Japan's typhoon approaches with custom probability forecasts. Its Agentic Delphi panel lets 2-3 AI evaluators assess evidence, challenge reasoning and revise estimates over two rounds. Track forecasts from initial judgment to scored outcomes. Conflict recurrence forecasting is in development. Philotica.com is the research hub connecting my ideas with the methods, tools and results of putting them into practice.

Add a comment

Replies

Best
Maker
📌
Hi everyone, I’m Mariusz, the creator of Philotica Forecast Studio. How would you estimate the probability of an event that matters to you, explain that estimate, and later check how well it held up? I’m building the Studio to connect those steps: historical data, statistical models, structured judgment, and evaluation against outcomes. Two statistical pilots are already working. One estimates probabilities for US natural gas storage using Energy Information Administration data. The other covers typhoon approaches to Japan using Japan Meteorological Agency records. You choose the parameters, inspect the method, and review historical performance against a baseline. For questions beyond those pilots, My forecast lets you define an event, deadline, evidence, and resolution criteria, then record and revise your probability estimate. The feature I’m particularly keen for you to try is the Agentic Delphi panel. Assembling a human expert panel takes time and access to people who may be unavailable. Agentic Delphi offers an accessible AI-based alternative for structured review, inspired by the Delphi method. Two or three evaluators, using models from OpenAI, Anthropic, or Google, assess your question separately. Their reasoning is challenged, a moderator summarizes their differences, and a second round lets them reconsider their estimates. You can inspect the reasoning and disagreement before accepting the result as a new version of your forecast. The panel can also review questions from the statistical pilots. Manual sessions use your own AI chats; automatic sessions use your API keys, with usage billed by the providers. This is an experimental beta. Agreement between models does not establish accuracy; that needs to be measured against actual outcomes. Next comes a module for armed conflict recurrence, using UCDP historical data. Alongside its statistical baseline, I plan to test whether hypotheses from my War-Peace Code research, including proposed cyclicality, add measurable predictive value. Philotica is the wider research hub where I publish those ideas, their statistical tests, interactive findings, and practical tools: I’m developing the project independently and welcome testers, research collaborators, and conversations with potential partners and investors. What question would you bring to the Studio, and what would you need to see before using its estimate in a real decision?