PRDs I've written

There's no universal PRD template, so I size each doc to its audience and stage. Every one still answers the same questions: what problem, who wants it and what's the evidence, how it fits, why now, what's critical, and how we'll know it worked.

    Automated Trading for NBA Playoff Win-Probability Markets

    Type
    Build spec
    Status
    Built for the 2026 playoffs
    Author
    @Alex

    At a glance: Kalshi's NBA playoff win-probability markets are priced by a crowd that overreacts, lags the court, and drifts from professional models, faster than anyone can trade by hand. This spec defines a bot that compares live ESPN game data with Kalshi odds, trades when a gap clears a threshold, and tags every trade with the hypothesis behind it, with hard risk limits that run independently of the strategy.

    Problem

    Kalshi's NBA playoff win-probability markets are priced by a crowd. Odds overreact to big moments, lag what's happening on the court, and drift from professional models, faster than anyone can trade manually.

    Three observed gap types, each treated as a separate hypothesis: real-time sentiment overreactions, the lag between the court and the TV broadcast, and differences between crowd odds and quant models.

    Current State

    Trading today: the gap closes before a human can act

    1. Big moment on court
    2. Crowd odds overreact or lag
    3. Trader spots the gap
    4. Gap closes before the trade
    manual trading · 4 steps, too slow for the window

    Proposed Approach

    A bot that compares live ESPN game data with Kalshi odds, trades when a gap clears a threshold, and logs which hypothesis triggered each trade. It reuses the modular pipeline from my Kalshi crypto bot: separate ingestion, comparison, and execution modules. Both feed a shared dashboard later.

    Same markets, acting in seconds instead of minutes

    Today
    1. Watch the game
    2. Check Kalshi odds
    3. Trade by hand
    Proposed
    1. Ingest ESPN + Kalshi
    2. Score each gap type
    3. Trade past threshold
    4. Log the hypothesis
    before/after · 3 steps today, 4 steps proposed

    What it does, by module

    • Ingest. Pulls live game state from ESPN and market odds from Kalshi, and flags stale data.
    • Detect. Scores each market for all three gap types, with a configurable threshold.
    • Execute. Places and manages contracts through the Kalshi API, with a max position size, a daily loss limit, and a manual kill switch.
    • Monitor. Logs every decision with its inputs and gap type. Next: a dashboard of open positions and P&L by gap type.

    Why now

    The playoffs are a fixed window. The bot had to be live for the first round or wait a year.

    Scope

    In scope

    • NBA playoff win-probability markets on Kalshi
    • All three gap types, each tracked as its own hypothesis
    • Automated execution through the Kalshi API
    • Risk limits and a kill switch that run independently of the strategy
    • Decision logging for every trade

    Out of scope

    • Regular season or other sports
    • Training my own prediction model

    Success Metrics

    The only reader is me, as builder and operator, so the bar is safe operation first and learning second.

    GoalMetricTarget
    Run safely unattended Unintended orders during a full playoff night Zero
    Make every trade explainable Trades traceable to the gap type that triggered them All of them
    Learn which hypotheses hold P&L broken out by gap type Reported per gap type

    Risks & Open Questions

    Risks

    • Stale ESPN or Kalshi data could trigger trades on a gap that isn't real, which is why ingestion flags stale data.
    • A bad strategy could lose money fast while unattended; position size, daily loss limit, and the kill switch cap that.
    • Some gaps may just be noise, so a threshold set too low would trade on randomness.

    Open questions

    • Which of the three gap types actually produces profitable trades?
    • Where should the threshold sit for each gap type?
    • How much of the crypto bot's pipeline carries over cleanly to the shared dashboard?