Multi-Source LLM Estimator

A prediction market trading bot that enriches an LLM with real-time context
from 10+ data sources. It compares the LLM's estimated probability against
the current market price and trades when it detects significant mispricing.

How It Works

CODEBLOCK0

1. 1. Fetches active markets from Polymarket (via Simmer SDK)
2. For each market, gathers real-time context from relevant data sources
3. Also checks cross-platform prices on Manifold and Kalshi
4. Sends the question + context to an LLM for probability estimation
5. If the LLM's estimate diverges from the market price by more than the

threshold (default 15%), it places a trade

Data Sources

#	Source	What It Provides	API Key Required
1	RSS News	Headlines from 10 feeds (Reuters, BBC, Bloomberg, etc.)	No
2

FRED | Economic indicators (Fed rate, CPI, GDP, VIX, oil, gold) | Yes (free) | | 3 | GDELT | Geopolitical event sentiment scores | No | | 4 | Odds API | Sports bookmaker consensus probabilities | Yes (free tier) | | 5 | FiveThirtyEight | US presidential approval polling averages | No | | 6 | Congress.gov | Bill status and legislative tracking | Yes (free) | | 7 | OpenFDA | Drug approval status and clinical trial data | No | | 8 | Open-Meteo | Weather forecasts for major cities | No | | 9 | USGS | Significant earthquake data | No | | 10 | Finnhub | Earnings calendar and IPO data | Yes (free) | | 11 | Manifold | Cross-platform prediction market prices | No | | 12 | Kalshi | Cross-platform prediction market prices | No |

Sources are selected automatically based on category detection from the
question text. Only relevant sources are queried to minimize latency.

Remixable Template

This skill is designed as a template you can customize:

• - Add your own data source: Implement a _get_X_context(question) function

that returns a list[str] of context lines. Add it to _dispatch_sources().

• - Swap the LLM: Set LLM_API_URL and LLM_MODEL env vars to point at

any OpenAI-compatible API (OpenRouter, Ollama, vLLM, etc.).

• - Adjust the threshold: Set ESTIMATOR_THRESHOLD higher for fewer but

higher-conviction trades, or lower for more frequent trading.

• - Change the prompt: Edit the _build_prompt() function to customize

the LLM's reasoning style.

Environment Variables

Required

Variable	Description
INLINECODE7	Simmer SDK API key for trading
INLINECODE8

API key for your LLM provider (default: OpenRouter) |

Optional

Variable	Default	Description
INLINECODE9	INLINECODE10	Trading venue (sim for paper, polymarket for real)
INLINECODE13

10.0 | Trade size in USD per trade | | LLM_API_URL | OpenRouter endpoint | OpenAI-compatible chat completions URL | | LLM_MODEL | xiaomi/mimo-v2-flash:free | Model identifier | | ESTIMATOR_THRESHOLD | 0.15 | Min divergence to trigger a trade (0.0-1.0) | | FRED_API_KEY | — | FRED API key (free at api.stlouisfed.org) | | ODDS_API_KEY | — | The Odds API key (free at the-odds-api.com) | | CONGRESS_API_KEY | — | Congress.gov API key (free) | | FINNHUB_API_KEY | — | Finnhub API key (free tier) | | LLM_CACHE_TTL | 1800 | LLM response cache TTL in seconds | | LLM_MAX_CALLS | 50 | Max LLM calls per run |

Usage

Dry-run (no trades, just log estimates):
CODEBLOCK1

Live trading:
CODEBLOCK2

Quiet mode (errors only):
CODEBLOCK3

Limit markets scanned:
CODEBLOCK4

Scheduling

Runs every 5 minutes via cron (*/5 * * * *). Managed automaton (auto-executes on schedule).
Capped at 50 LLM calls per run (LLM_MAX_CALLS) to control costs.

Security

• - All trades go through SimmerClient.trade() only. No direct CLOB or wallet access.
• Dry-run by default. The --live flag must be explicitly passed to execute trades.
• No wallet private keys are required or read by this script.
• LLM receives only the market question text and publicly available context (news headlines, economic indicators, odds). No credentials or private data are sent to the LLM.
• INLINECODE32 defaults to OpenRouter. You control which LLM endpoint is used.
• All optional API keys (FRED, Odds, Congress, Finnhub) are for free public data APIs. If unset, those data sources are simply skipped.