Kalshalyst — Contrarian Prediction Market Scanner

Overview

Kalshalyst is a complete intelligence system for finding and trading prediction market opportunities. It combines:

• - Claude Sonnet contrarian estimation — sees market prices and finds reasons they're WRONG
• Brier score tracking — measures how well your estimates calibrate against actual outcomes
• Kelly Criterion position sizing — calculates optimal trade size for each opportunity
• Five-phase pipeline — FETCH → CLASSIFY → ESTIMATE → EDGE → ALERT

The key insight: blind estimation (not seeing prices) produces consensus-matching estimates with zero edge. Contrarian mode (showing Claude the price and asking it to disagree) produces opinionated, directional calls with real edge.

When to Use This Skill

• - You want to find mispricings on Kalshi prediction markets
• You're looking for contrarian opportunities where the market is wrong
• You need to track how accurate your probability estimates are over time
• You want to size positions intelligently based on edge and confidence
• You're building a systematic prediction market trading system

First run does not require Kalshi credentials. If they are missing, Kalshalyst prints a realistic demo scan and writes demo cache data so downstream tools like Market Morning Brief still have something useful to show.

Requirements

API Keys & Credentials

1. 1. Kalshi API Key (free at kalshi.com)

- Sign up at https://kalshi.com - Navigate to Settings → API - Generate API credentials (key ID + private key file) - Cost: Free tier supports unlimited reads, small position limits

1. 2. Anthropic API Key (for Claude Sonnet)

- Create account at https://console.anthropic.com - Generate an API key - The public reference implementation calls Anthropic directly - Budget: variable by scan volume; expect non-zero Claude cost if you run frequent scans

1. 3. Polygon.io API Key (optional, free tier available)

- Sign up at https://polygon.io - Free tier includes market data + basic news - Cost: Free tier sufficient, paid plans for higher volume

Python & Dependencies

• - Python 3.10 or higher
• Required packages:

CODEBLOCK0

• - Optional (for local fallback estimation):

- Ollama (https://ollama.ai) with Qwen model - Install from https://ollama.ai (macOS, Linux, Windows) - Then: INLINECODE0

Configuration

Create or update your config.yaml file with:

CODEBLOCK1

The Five-Phase Pipeline

Phase 1: FETCH — Kalshi Market Discovery

Fetches all open markets from Kalshi and applies pre-filters:

Blocklist Filtering:

• - Ticker prefixes: KXHIGH, KXLOW, KXRAIN, KXSNOW, INX, NASDAQ (weather, intraday noise)
• Category slugs: weather, climate, entertainment, sports, social-media
• Micro-timeframes: "in next 15 min", "in next 5 hours" (coin flips)
• Sports tokens: NFL, NBA, soccer, esports (blocked from the production stack)

Timeframe Gates:

• - Minimum days to close: 7 (default)
• Maximum days to close: 365 (default)
• Markets without expiration dates are blocked (usually garbage)

Volume Floor:

• - Minimum volume: 50 contracts (default)
• Filters out illiquid, low-interest markets

Price Availability:

• - Must have at least one price signal (yesbid, yesask, yesprice, or lastprice)
• Resolves multiple price sources (bid/ask mid preferred)

Output: List of ~100-500 pre-filtered markets ready for analysis

Phase 2: CLASSIFY — Market Classification

Status: Disabled (Qwen unreliable) — Markets pass through with defaults.

When re-enabled, would use local Qwen to classify each market by:

• - Category: politics, economics, crypto, policy, technology, etc.
• Tradability: 0.0-1.0 score (how analyzable with public info?)
• News sensitivity: True if breaking news would materially shift the probability

For now, all markets receive default classification values and proceed to Phase 3.

Phase 3: ESTIMATE — Claude Contrarian Probability Estimation

The core IP. Claude sees the market price and is asked to find reasons it's WRONG.

System Prompt (Contrarian Mode):
CODEBLOCK2

Context Enrichment:

• - Recent news from Polygon.io (if configured)
• Macro indicators: S&P 500, Bitcoin, VIX proxy, Gold
• X/Twitter sentiment signals (if available)
• Market liquidity and volume

Fallback to Qwen:

• - If Claude is unavailable (cooldown, network error), falls back to local Qwen
• Qwen runs blind (no price shown) to prevent anchoring
• Falls back to full Qwen batch mode after 3 consecutive Claude failures

Output: Estimated probability, confidence, reasoning, key factors

Phase 4: EDGE — Edge Calculation

For each estimate, calculates the edge (profit potential):

CODEBLOCK3

Why limit orders? You're a sophisticated trader who can post limit orders at midpoint or better, avoiding spread costs.

Filtering:

• - Minimum effective edge: 3% (default, configurable)
• Minimum confidence: 0.4 (filter out uncertain estimates)
• Drop estimates with direction = "fair" (no edge)

Output: Ranked list of edges, highest first

Phase 5: CACHE & ALERT

Cache Writing:

• - Writes research cache to .kalshi_research_cache.json (used by related commands)
• Detailed edge data to state/kalshalyst_results.json for analysis

Alerting:

• - Filters opportunities by alert threshold (default: 6% effective edge)
• Sends top 3 opportunities to the user
• Format: ticker, direction (YES/NO), probability, edge, confidence, reasoning

Brier Tracking:

• - Every estimate is logged to SQLite database
• Computes info_density (news + X signals + economic context + liquidity)
• Used for later calibration analysis

Blocklist System

Complete Blocklist Reference

Ticker Prefixes (High-Volume Garbage):
CODEBLOCK4

Category Slugs (API-Level Blocking):
CODEBLOCK5

Micro-Timeframe Patterns (Coin Flips):
CODEBLOCK6

Sports Tokens (Blocked From The Production Stack):

• - Major leagues: NFL, NBA, MLB, NHL, MLS, NCAA, PGA, UFC, WWE
• Soccer: Premier League, La Liga, Serie A, Bundesliga, Champions League, Copa
• Esports: Valorant, League of Legends, CS:GO, Dota, Overwatch
• Individual sports: ATP, WTA, Tennis, Boxing, MMA

Why These Filters?

• - Weather + Intraday: Near-pure noise — impossible to extract edge
• Sports: Intentionally excluded. Recent evaluation did not show durable model edge, so sports are not part of the current production stack.
• Entertainment: Celebrity/social media volatility — not analyzable with Claude
• Micro-timeframe: Spreads dominate, zero informational edge
• Blocklist philosophy: Cut the bottom 80% of opportunities (noise) to focus Claude on the top 20% (signal)

Contrarian Estimation — Why It Works

The Problem with Blind Estimation

Blind mode (not showing Claude the market price):

• - Claude produces "consensus" estimates
• Usually close to 50% for uncertain markets
• Results in zero edge (estimate ≈ market price)
• Not actionable

Why? Claude doesn't know what the market thinks, so it defaults to high uncertainty.

The Solution: Contrarian Mode

Contrarian mode (showing Claude the price):

• - Claude sees the market price: "Market is priced at 35%"
• Prompt asks: "Is this price WRONG?"
• Claude identifies reasons for disagreement:

- Missing recent news
- Crowd psychology error
- Timing mismatch
- Base rate neglect

• - Result: Opinionated directional call (e.g., 62% vs 35% market) → 27% edge

Example

Market: "Will Ukraine still be at war in 2026?"
Market Price: 72% (market implies YES very likely)
Recent Context: Leaked peace negotiations, US pushing settlement

Claude Contrarian Reasoning:
CODEBLOCK7

Edge: |38% - 72%| = 34% effective edge → BUY NO at 28¢ expected return

Brier Score Tracking

What It Measures

Brier Score = (1/n) * Σ(forecast - outcome)²

• - 0.0 = perfect estimates
• 0.25 = random baseline (coin flip for 50/50 events)
• Above 0.25 = worse than guessing (miscalibrated)

How It Works

1. 1. Log Phase: Every edge scanner run logs all estimates to SQLite

- Ticker, estimated probability, market price, confidence, estimator (Claude vs Qwen) - Category (politics, policy, crypto, etc.) - Edge percentage, info density (context richness score)

1. 2. Resolve Phase: When markets close on Kalshi, log the outcome

- Automatic: check_and_resolve_markets() polls Kalshi API daily - Manual: INLINECODE5

1. 3. Report Phase: get_brier_report() computes calibration

- Overall Brier score - Breakdown by estimator (Claude vs Qwen accuracy) - Breakdown by category (politics, crypto, policy) - Calibration buckets: when you say 70%, does it resolve YES ~70%? - Win rate for "edge trades" (edge >= 4%)

Database Schema

CODEBLOCK8

Info Density Scoring

Measures how much context was available at estimation time (0.0-1.0):

• - News articles (0.0-0.25): 0-3 Polygon articles → 0/0.15/0.25
• X signals (0.0-0.25): Corroborating social signal from X scanner
• Economic context (0.0-0.25): S&P 500, Bitcoin, VIX available
• Liquidity (0.0-0.25): Volume + open interest proxy

Higher density estimates should be better calibrated (more informed).

Calibration Alerts

Weekly check: get_calibration_alert()

• - Identifies categories with Brier > 0.25 (worse than random)
• Suggests recalibrating estimator prompts for those categories
• Example: "Politics: Brier 0.31 (5 resolved) — systematically overconfident"

Kelly Criterion Position Sizing

The Math

For binary prediction markets, Kelly fraction tells you what % of bankroll to risk:

CODEBLOCK9

Example: Estimate YES at 65%, market prices at 50¢
CODEBLOCK10

With $200 bankroll → $60 bet = 120 contracts at 50¢

Fractional Kelly (Conservative)

Full Kelly is risky with noisy estimates. Kalshalyst uses fractional Kelly (α = 0.25 default):

CODEBLOCK11

Same example, confidence = 0.7:
CODEBLOCK12

Hard Caps (Defense-in-Depth)

• - Max contracts per trade: 100
• Max cost per trade: $25
• Max portfolio exposure: $100 (user-configurable)
• Minimum edge: 3% (below this, noise dominates)

Configuration

CODEBLOCK13

Output Format

CODEBLOCK14

Scheduling & Deployment

Command-Line Usage

CODEBLOCK15

As a Cron Job (Every 60 Minutes)

CODEBLOCK16

As a OpenClaw Scheduled Task

CODEBLOCK17

Docker Deployment

CODEBLOCK18

Example Output

Alert Message

CODEBLOCK19

Research Cache (Programmatic Access)

CODEBLOCK20

API Reference

Main Functions

CODEBLOCK21

Troubleshooting

Claude Unavailable (Rate Limited / Cooldown)

Kalshalyst automatically falls back to Qwen after 3 consecutive Claude failures. Check:

• - Anthropic API key validity and quota
• Network connectivity
• Rate limits (Claude has usage-based limits)

To force Qwen fallback immediately:
CODEBLOCK22

No Markets Passing Filters

If "no markets passed filters":

• - Kalshi API may be down (check https://status.kalshi.com)
• All markets may be blocklisted (verify blocklist config)
• Check your network — fetch may have timed out

To debug:
CODEBLOCK23

Brier Score Not Computing

Brier score requires market resolutions. If none available:

1. 1. Wait for markets to close (usually 24-48 hours)
2. Manually log resolutions: INLINECODE8
3. Check database: INLINECODE9

Position Sizing Returns 0 Contracts

Possible causes:

• - Edge below 3% minimum (check min_edge_for_sizing)
• Confidence below 0.2 threshold
• Bankroll too low or exposure at limit
• Kelly fraction negative (bad odds for the bet)

Verify with:
CODEBLOCK24

Performance & Cost

Typical Run Metrics

• - Runtime: 2-4 minutes (50 markets, Claude estimation)
• API calls:

- Kalshi: 1-10 calls (market fetch, pagination) - Claude: 50-80 calls (estimate_batch) - Polygon: 2 calls (economic indicators, every 12 hours)

• - Cost per run:

- Claude: variable by model and usage volume - Polygon: ~$0 (free tier) - Kalshi: $0 (read-only)

Scaling

For scheduled operation, Claude spend scales directly with your scan frequency and model selection. If you want a zero-API-cost fallback, keep Ollama/Qwen available and treat it as a lower-quality backup path rather than the primary estimator.

OpenClaw Ecosystem Integration

Kalshalyst is the intelligence engine of the Prediction Market Trading Stack. It feeds edge data to other skills:

Connected Skill	What It Gets From Kalshalyst
Market Morning Brief	Top edges appear in your daily morning digest
Kalshi Command Center

Execute trades directly on flagged opportunities |
| Portfolio Drift Monitor | Track positions opened on Kalshalyst recommendations |

Install the complete stack:
CODEBLOCK25

Implementation Notes

Battle-tested in production trading environments. Key design decisions:

1. 1. Generic config loading — works with any OpenClaw config.yaml
2. Direct alerts — sends opportunities directly to you, no middleware
3. All thresholds, blocklists, and prompts refined through live trading
4. Scripts are standalone — zero external dependencies beyond listed packages

Further Reading

• - See references/contrarian-prompt.md for the full Claude system prompt with explanation
• See references/blocklist.md for complete blocklist documentation
• See references/brier-schema.md for database schema and methodology
• See references/kelly-math.md for Kelly Criterion math derivation

Support & Iteration

Kalshalyst is actively maintained. Common iteration paths:

1. 1. Contrarian Prompt Tuning: Refine system prompt to improve edge quality
2. Blocklist Refinement: Add new ticker prefixes or patterns as you encounter noise
3. Category Recalibration: Use Brier reports to identify systematically miscalibrated categories
4. Kelly Parameters: Adjust alpha, max contracts, exposure limits based on risk tolerance
5. Fallback Estimators: Replace Qwen with custom local models or other APIs

See the GitHub issues (if published) or contact the maintainer for feature requests.

---

Agent Bug-Fix Protocol

CRITICAL: When you (the agent) discover and fix a bug in any script during a conversation, you MUST write the fix to the source file on disk using file editing tools. Showing the fix in chat is NOT sufficient — the next scheduled run starts a fresh session and will not have your conversational fix.

Checklist after finding a bug:

1. 1. Edit the actual .py file on disk (e.g., ~/skills/kalshalyst/scripts/*.py)
2. Also update the installed copy at INLINECODE17
3. Confirm both files are written before telling the user the fix is done
4. If the fix is significant, stage a git commit in INLINECODE18

Never say "fixed" if you only demonstrated the fix in conversation output.

Feedback & Issues

Found a bug? Have a feature request? Want to share results?

• - GitHub Issues: github.com/kingmadellc/openclaw-prediction-stack/issues
• X/Twitter: @KingMadeLLC

Part of the OpenClaw Prediction Stack — the first prediction market skill suite on ClawHub.