An open-source bot that trades prediction markets on Kalshi. It scans markets, enriches them with real-time data (economic indicators, weather, sports odds), gets independent analyses from multiple AI models, and trades when a consensus forms with sufficient edge.
Maybe. This is experimental software. Run in dry-run mode (the default) until you have 50+ resolved trades with a positive win rate. Even then, prediction markets are inherently uncertain. Never risk money you can't afford to lose.
A Kalshi account with API access and at least one LLM API key (Anthropic, OpenAI, or xAI). A local LLM via LM Studio is recommended for free screening -- it saves roughly $4/day in cloud API costs. A FRED API key (free) significantly improves economic market predictions.
With a local LLM handling the initial screen, cloud API costs run $1-3/day depending on market volume. Without a local LLM, expect $5-8/day. Infrastructure costs are negligible -- it runs on any machine with Python 3.10+.
No minimum, but the bot defaults to small position sizes. The default max bet is $25. You can adjust MAX_BET_AMOUNT and KELLY_FRACTION in your .env file.
No single model is reliably calibrated. By having 3-4 models analyze each market independently, the system filters out individual model biases. A trade only happens when a majority agrees on direction and their probability estimates are within 20% of each other.
After each market settles, the bot records whether it won or lost. Brier scores, win rates by category and confidence level, and the reasoning from winning and losing trades are all injected into future LLM prompts. The system learns from its own history.
Kalshi markets pay $1 or $0 at settlement. Mid-market price fluctuations are noise. Stop-losses that exit at, say, $0.30 on a position entered at $0.50 lock in a loss that might have settled at $1. Backtesting across multiple strategies confirmed that holding to settlement outperforms active exit management.
yfinance for market data (S&P 500, Nasdaq, commodities, crypto, forex). FRED for economic indicators (payrolls, CPI, unemployment, gas prices). NWS for weather forecasts across 21 US cities. ESPN for live sports odds. All woven into the context the LLMs see when analyzing a market.
An opt-in network where ktrader instances share anonymized predictions. Each bot submits its probability estimate, direction, and confidence for every market it analyzes. The server aggregates these signals and returns the crowd view to all members.
The collective's edge is inversely proportional to its size. With hundreds of members, each gets a meaningful signal -- independent bots with different models, data, and strategies all looking at the same market. At mass adoption, the collective signal converges to the market price itself. The cap keeps it useful.
Market ticker, your bot's probability estimate, side (yes/no), confidence level, model source, and the eventual outcome. That's it. No entry price, no P&L, no balance, no reasoning text, no API keys, no personal information.
Yes -- the server code is MIT licensed and included in the repo. But a collective with 3 members produces noise, not signal. The value comes from the network: hundreds of independent bots analyzing the same markets. A fork starts at zero participants and zero history.
It's entirely optional. The bot works identically without it. Set COLLECTIVE_ENABLED=false (the default) and it never contacts the collective server.
No. Your Kalshi API key and private key never leave your machine. They're used locally to sign API requests to Kalshi directly. The collective uses a separate key that you choose yourself.
Your identity in the collective is a truncated SHA-256 hash of the key you choose at signup. There is no account creation, no email required to join, and no way to link your collective activity to your Kalshi account.
It starts in dry-run mode by default and logs paper trades without placing real orders. Circuit breakers halt trading if daily losses exceed a configurable threshold. That said, this is autonomous trading software -- review the code, understand the risks, and start small.