Framework

XRTM is built on a 4-layer framework. Each layer is an independent Python package. Use them directly to build custom forecasting systems.

Packages

Package	Version	Role	Install
xrtm-data	0.3.0	Schemas & question sources	pip install xrtm-data
xrtm-eval	0.3.0	Scoring (Brier, ECE, log)	pip install xrtm-eval
xrtm-forecast	0.7.0	Runtime engine	pip install xrtm-forecast
xrtm-train	0.3.0	Backtesting & optimization	pip install xrtm-train

Quick Examples

Load Questions

from xrtm.data.corpora import load_real_binary_questions
questions = load_real_binary_questions(limit=5)

from xrtm.data.providers.online import PolymarketSource
markets = await PolymarketSource().fetch_questions(limit=10)

Score Forecasts

from xrtm.eval import BrierScoreEvaluator, summarize_binary_forecasts

evaluator = BrierScoreEvaluator()
brier = evaluator.score(probability=0.7, ground_truth="yes")

predictions = [(0.7, "yes"), (0.3, "no"), (0.9, "yes")]
summary = summarize_binary_forecasts(predictions)
# {'brier_score': 0.0467, 'ece': 0.1333, 'log_score': 0.1625}

Run Forecasts

from xrtm.forecast.kit.agents.specialists.analyst import ForecastingAnalyst
from xrtm.forecast.core.config.inference import OpenAIConfig
from xrtm.forecast.providers.inference.factory import ModelFactory

config = OpenAIConfig(
    model_id="your-model",
    base_url="https://your-endpoint",
    api_key="sk-...",
)
provider = ModelFactory.get_provider(config)
analyst = ForecastingAnalyst(model=provider)

from xrtm.data.corpora import load_real_binary_questions
question = load_real_binary_questions(limit=1)[0]
forecast = await analyst.run(question)

Provider Support

OpenAI-compatible only. Any endpoint that speaks the OpenAI Chat Completions API works.

Agents

• ForecastingAnalyst — Structured probabilistic forecasts with causal reasoning traces
• LLMAgent — Provider-backed agent with tool retrieval
• RoutingAgent — FAST/SMART tiered routing

Topologies

Composable multi-agent patterns:

• RecursiveConsensus — Parallel analysts → aggregate → supervisor loop
• create_debate_graph() — Pro/Con/Judge debate
• create_fanout_graph() — Fan-out to parallel workers

Backtesting

from xrtm.train import Backtester, BacktestRunner

Dependencies

Each package is intentionally lean:

• xrtm-data: pydantic, scipy
• xrtm-eval: pydantic
• xrtm-forecast: pydantic, openai, aiohttp, sqlalchemy (cache), numpy, freezegun, rich
• xrtm-train: pydantic + the three framework packages above

Architecture →