Forecast Conditional Scenarios

A single probability often hides the interesting part. "When will the ban be lifted?" gets a wide, mushy distribution, because the answer depends almost entirely on why the ban happened in the first place. Split it by cause and each branch becomes tight and defensible.

That is what a conditional forecast does. You supply a condition, and the outcome is forecast twice: once in the world where the condition holds, once in the world where it does not. Both branches are forecast jointly, so they stay coherent with each other rather than drifting apart the way two separate runs would.

Conditional forecasting is a modifier, not a forecast type. It stacks on top of binary, numeric, date, categorical or thresholded. The forecast_type still describes the outcome; the condition describes the world it happens in.

Two ways to supply the condition

The two are mutually exclusive. Conditional forecasts require effort_level="HIGH", which is also the default.

See it in production

We forecast the same date under three different assumed causes, alongside the unconditional forecast:

• Restoration date, unconditional
• Restoration date, assuming a sincere but mistaken determination
• Restoration date, assuming a genuine foreign access risk

The unconditional forecast is a blend of the branches. On its own it tells you less than the branches do, which is the whole argument for running the conditional version.

OpenAI post-IPO market cap, conditional on a step-change model shipping first shows the same idea on a numeric question.

claude mcp add futuresearch --scope project --transport http https://mcp.futuresearch.ai/mcp

Forecast when the US restriction on this model will be lifted,
conditional on the restriction having been driven by a genuine
foreign access risk rather than a policy error.

Tool: futuresearch_forecast
├─ data: [{"question": "When will the US restriction be lifted?"}]
├─ forecast_type: "date"
├─ output_field: "restoration_date"
├─ condition: "The restriction was driven by a genuine foreign access risk"
└─ effort_level: "HIGH"

→ Submitted: 1 row, conditional date forecast.

pip install futuresearch
export FUTURESEARCH_API_KEY=your_key_here  # Get one at futuresearch.ai/app/api-key

import asyncio
import pandas as pd
from futuresearch.ops import forecast

labs = pd.DataFrame([
    {"question": "What will Lab A's 2027 revenue be, in billions USD?"},
    {"question": "What will Lab B's 2027 revenue be, in billions USD?"},
    {"question": "What will Lab C's 2027 revenue be, in billions USD?"},
])

async def main():
    result = await forecast(
        input=labs,
        forecast_type="numeric",
        output_field="revenue",
        units="billions USD",
        condition="US compute export controls tighten further during 2026",
        effort_level="HIGH",
    )
    print(result.data[[
        "question",
        "revenue_p50_given_condition",
        "revenue_p50_given_not_condition",
    ]])

asyncio.run(main())

scenarios = pd.DataFrame([
    {
        "question": "When will the restriction be lifted?",
        "cause": "The restriction was a sincere but mistaken determination",
    },
    {
        "question": "When will the restriction be lifted?",
        "cause": "The restriction was driven by a genuine foreign access risk",
    },
])

result = await forecast(
    input=scenarios,
    forecast_type="date",
    output_field="restoration_date",
    condition_field="cause",
    effort_level="HIGH",
)

Reading the output

A conditional forecast replaces the normal forecast columns with a _given_condition and a _given_not_condition copy, plus one shared rationale covering both branches.

If the two branches come back nearly identical, the condition does not actually bear on the outcome. That is a real finding, and it is worth more than a confident single number.

When to use decision() instead

Conditional forecasting answers "what happens if the world turns out this way". If the condition is something you control, such as how much to fund a project or which vendor to pick, use decision(). It forecasts the outcome under each alternative of the choice, under explicit intervention assumptions, which is a causal question rather than a correlational one.

Built with FutureSearch. See the forecast documentation for all parameters and output formats. Related guides: Forecast Categorical and Threshold Questions, Turn Claude into an Accurate Forecaster, Forecast Outcomes for a List of Entities.