Lightning Rod Python SDK

Foresight returns a calibrated probability for any question about the future — through an OpenAI-compatible API. No fine-tuning, no setup. Ranked #1 for forecasting accuracy on ProphetArena.

Trusted for high-stakes predictions by Numinous, Shore Capital Partners, Awardable (Tradewinds Solutions Marketplace), and ERIS Marketplace. Foresight processes billions of tokens and serves 100k+ inference requests every day.

Documentation · Get an API key · Research paper

⚡ Better predictions in seconds

Foresight is served behind an OpenAI-compatible endpoint, so any OpenAI client works — just point base_url at Lightning Rod.

pip install openai

from openai import OpenAI

client = OpenAI(
    base_url="https://api.lightningrod.ai/api/public/v1/openai",
    api_key="your-api-key",
)

completion = client.chat.completions.create(
    model="LightningRodLabs/foresight-v4",
    messages=[
        {"role": "user", "content": "Will the Fed cut rates at its next meeting?"},
    ],
    extra_body={
        "research": True,       # gather live web evidence before forecasting
        "answer_type": "auto",  # append a structured answer in <answer></answer> tags
    },
)

print(completion.choices[0].message.content)
# Foresight weighs the evidence and returns a calibrated probability,
# e.g.: "... <answer>0.34</answer>"

That 0.34 is a calibrated probability — a 34% chance, not a confidence score or a yes/no. 0.5 means genuinely uncertain, and across many ~0.7 forecasts roughly 70% should come true. See the forecasting reference for answer types, research sources, and the full response shape.

Prefer an SDK helper?

lr.predict() wraps the same API and parses the structured answer for you:

pip install lightningrod-ai openai

import lightningrod as lr

client = lr.LightningRod(api_key="your-api-key")
result = client.predict(
    "Will the Fed cut rates by 25bp in March 2026?",
    answer_type="binary",
    research=True,
)
print(result.binary.probability)  # e.g. 0.62

🏗️ Train an expert on your domain

Need a model tuned to your domain? Our enterprise platform enables companies to turn raw sources into labeled datasets and fine-tunes models on them — served through the same API.

📅 Book a call with us

pipeline = QuestionPipeline(...)
dataset = client.transforms.run(pipeline)

train_dataset, test_dataset = prepare_for_training(dataset)
train_config = GRPOTrainingConfig(base_model_id="openai/gpt-oss-120b")
training_job = client.training.run(train_config, train_dataset)

# your fine-tuned model is served through the same predict() / OpenAI API
client.predict("Will the Fed cut rates by 25bp in the next 3 months?", model=training_job.model_id)

We used this to generate the Future-as-Label training dataset from our paper, Future-as-Label: Scalable Supervision from Real-World Outcomes.

📚 Learn more

• Documentation — quickstart, guides, recipes, and the REST API reference
• Example notebooks — forecasting, dataset generation, training, and evaluation (runnable in Colab)
• SDK API reference — every class and method in this repo