Scientific Forecasting Company (SFC)
Predicting future world events to better understand what to do now.
Using smaller open-source models fine-tuned with LoRA, we aim to give people, institutions, and governments accurate predictions that benefit everyone.
Example forecast
Will Putin hold a bilateral meeting with Alexander Lukashenko in Turkmenistan?
Our forecast (Oct 10, 2026): 2%
Prediction market at the time: 42%
Result: No
One of our live forecasts.
How SFC forecasts
Beyond the headlines.
A comprehensive historical database we call "the brain" lets the model research far deeper than standard prediction bots. Next, the model will be able to ask itself questions and find its own answers in the brain, mimicking how humans forecast while weighing thousands more pieces of information than a human could.
The news.
Using AskNews, our model gathers roughly 20 relevant articles from trusted news sources, then reads, analyzes and weighs them before reaching a conclusion.
Our own model.
The model is trained by rewarding accuracy, measured with Brier scores, the industry standard for scoring predictions. It has already been trained and scored on over a thousand real questions.
At a glance
14B parameters in our model.
~10,000 questions to train on.
~1,000 H100 GPU hours of training.
~6 weeks of nonstop training on a single GPU.
~40,000 practice forecasts written and scored.
~200,000 news articles read during training.
Covers politics, economics, and world events.
Figures are estimates for our next training run.
Pricing
TBD
About
Noah Stecklov
I'm a final-year student at McMaster University, studying mathematics and economics. Forecasting bots caught my attention in early 2026, and the field seemed interesting and still not very saturated. I believe I can add something unique to the forecasting industry.