Foresight by Lightning Rod - Predict anything with AI
by•
Foresight by Lightning Rod is an OpenAI-compatible forecasting API for developers building agents, prediction-market bots, and decision tools. Ask a question about a future event and get a scored, calibrated forecast back. Unlike general-purpose LLMs, Foresight is trained and evaluated on real-world outcomes, with benchmark-verified accuracy, cheaper inference, and a drop-in API for forecasting workflows.
Replies
Earth.fm
Congrats on the launch! 🚀 Training data is still one of the biggest bottlenecks in building reliable AI systems, so it's great to see a solution focused on turning real-world data into high-quality, traceable datasets instead of relying on manual labeling. The provenance and quality-scoring features really stand out. Excited to see how teams use this for domain-specific fine-tuning. Wishing you an amazing launch! 👏
Ensembling to get bands around the probabilities is exactly the part I'd reach for, since in a decision loop a miscalibrated tail costs you asymmetrically more than a wrong point estimate. The tension I keep hitting: ensembling N models fights the cheap-inference pitch the moment an agent is forecasting thousands of markets a run. Do you expose the band per-call in the API response, or is it a heavier mode you opt into when the stakes justify the extra passes?
Lightning Rod: AI Forecasting API
@dipankar_sarkar Absolutely, plenty of our users will run ensembles– running the same forecast multiple times (even with slightly different context / prompts) to get a distribution of results. We do see this meaningfully improves accuracy in our testing.
This is a big reason why having an economical LLM for forecasting is so important–cheaper inference makes it possible to run everything you need to run at scale without eating too much of your margin.
how often do you recalibrate the model as new outcomes become available?
Lightning Rod: AI Forecasting API
@maryam_nafees1 We typically update the models every few months, but we actually haven't observed much drift. What we're really training is better reasoning over real-world outcomes, and the model learns to use whatever information is shared in context. So most of the gains don't require the model itself having fully up-to-date knowledge of the world, as long as the right info is in context.
OpenAI-compatible is a smart wedge — I can point an existing client at it and test forecasting without rewriting my stack. The thing I'd check first: when the model is genuinely uncertain, does the API return a calibrated probability or confidence interval I can threshold on, or just a point prediction I have to trust blindly? For anything I'd wire into a real workflow, knowing when to ignore the forecast matters more than the forecast itself.
Lightning Rod: AI Forecasting API
@hazy0 Thanks Hazy - we have built in "answer types" you can use to easily extract a prediction for different question types (e.g. binary if you just want a probability), but you can give the model instructions for any custom answer format also. So if you want a probability + confidence, or low/high estimates, you can ask for that also! I agree it is different if the model is uncertain because it doesn't have enough information vs being confident that the outcome is a toss-up. You can also do back-testing to determine what category of questions or kind of context the model performs well at and just apply it to those questions.
Another thing you can do if you are applying it to a question with a market price or established prior is include that in the prompt, and ask explicitly whether it wants to make an adjustment to that price or not.
CheckYa
Really like this idea! Clean, practical, and solves a challenge that almost every AI team faces. Looking forward to seeing how this evolves—best of luck with the launch! 🚀
Predicting anything is a bold promise, and I like that you're not boxing it into one niche like sports or markets. That open-ended framing is fun to play with.
The OpenAI-compatible API shape seems like a smart choice here, because forecasting is often something an agent wants to call inside a larger workflow, not a separate dashboard.
One product detail I’d look for as a developer is whether the calibration stays visible after the API call: confidence interval, data/source freshness, and a short reason why the model thinks the probability changed. For agent workflows, the forecast is useful, but knowing when not to trust it may be even more useful.
I like that you're tackling forecasting as its own problem instead of assuming a general-purpose LLM is the right tool for everything. An OpenAI-compatible API with lower inference costs also makes it much easier for developers to experiment without reworking their existing workflows. Beyond sports and politics, which real-world forecasting domains have you found Foresight performs especially well in, and are there any areas where you'd still recommend using a frontier model instead?
Macaly
calibrated forecast api is a sharp angle, way more usefull than asking a generic llm 🔮 whats the use case ppl reach for first?
how much better than just prompting claude. got any benchmarks