What Is This?

This is an automated scanner that identifies potentially mispriced daily high temperature contracts on Kalshi, a CFTC-regulated prediction market exchange. It compares probability estimates derived from National Weather Service forecasts against live Kalshi market prices to find contracts where the market may disagree with the NWS forecast.

How Kalshi Weather Markets Work

Kalshi offers daily high temperature markets for major US cities. Each market asks a question like "Highest temperature in NYC today?" and offers multiple brackets (e.g., ≤48°F, 49-50°F, 51-52°F, ≥57°F). Each bracket is a YES/NO contract that settles at $1.00 if the actual NWS-reported high falls in that range, or $0.00 if it does not.

Prices range from 1¢ to 99¢ and represent the market's implied probability. Settlement is based on the official NWS Daily Climate Report from that city's designated weather station.

Cities Tracked

The scanner monitors all active Kalshi weather markets. Each city settles on a specific NWS weather station:

City	NWS Station	Kalshi Series

The Probability Model

Core Concept

We model the actual daily high temperature as a normally distributed random variable centered on the NWS forecast high. The standard deviation is calibrated using historical forecast error data for each station. We then calculate the probability that the temperature falls within each Kalshi bracket by integrating the normal CDF across the bracket bounds.

Step 1: Forecast Mean

• NWS forecast high/low — from the official NWS point forecast. For HIGH markets, the daytime period high. For LOW markets, the minimum of the preceding night and following night periods (matching Kalshi's midnight-to-midnight settlement window).
• Auto-calibrated bias — an exponential moving average of recent forecast errors per city. Updated automatically after each settlement: new_bias = 0.3 × latest_error + 0.7 × old_bias. Activates after 3 settlements. Before that, a static default is used.
• Trend adjustment — if current observations show the day running hotter or colder than forecast, the mean is nudged. For HIGH markets, this applies during 8am-4pm ET weighted by time of day. For LOW markets, if the observed minimum is already below the forecast low, the mean shifts down aggressively (70-90% weight depending on whether temps are still falling or rising).

Step 2: Standard Deviation (Sigma)

Sigma represents how uncertain the model is. A larger sigma means wider probability spreads and fewer confident signals. The final sigma is computed from four stacked multipliers:

Base Sigma (Auto-Calibrated)

Instead of hardcoded values, the base sigma is learned from data:

• 90-day backtest — on startup, we pull 90 days of historical daily highs/lows from NOAA's RCC-ACIS climate database for each city. The standard deviation of day-to-day temperature changes gives us the baseline sigma.
• Rolling calibration — after each settlement, the rolling sigma is updated from the last 14 days of actual forecast errors. This adapts automatically — if a city has been hard to forecast recently, sigma widens. Activates after 5 settlements.
• Static fallback — until enough data accumulates, a per-city default from the original spec is used.

The CLI and dashboard show [cal] when calibrated values are active, or [static] when using defaults.

Confidence Multiplier

Confidence	Multiplier	When Used
High	×0.45	Day-of forecast, stable weather pattern
Medium	×0.70	2-3 days out, some uncertainty
Low	×1.0	4+ days out or unstable pattern

Weather Pattern Stability (×0.70 to ×1.50)

The scanner analyzes three real-time data sources to assess how stable the weather pattern is for each city:

Source	What It Measures	Range
Hourly Forecast Volatility	24-hour temperature range and max hour-to-hour swings from the NWS hourly forecast	×0.85 to ×1.30
Active Weather Alerts	NWS watches, warnings, and advisories for the city's coordinates, scored by severity	×1.0 to ×1.50
Forecast Discussion (AFD)	NWS meteorologist narrative analyzed for instability keywords ("cold front", "storm", "uncertainty") vs stability keywords ("high pressure", "ridge", "calm")	×0.85 to ×1.25

These three factors multiply together and are clamped to a 0.70–1.50 range. The dashboard shows: 🟢 very stable, 🟡 normal, or 🔴 volatile for each city.

Time-of-Day Decay (×0.40 to ×1.0)

As the day progresses and more observations come in, the model becomes more certain about the outcome. Sigma shrinks accordingly:

HIGH Markets	Decay	Reasoning
Before 8am ET	×1.0	Full uncertainty — high hasn't been reached
8-11am ET	×0.92	Morning — temps climbing
11am-2pm ET, near forecast	×0.70	Observed max approaching forecast
After 2pm, max ≈ forecast	×0.50	Nearly certain — high is set
After 2pm, max > forecast	×0.40	Already exceeded — outcome clear

LOW Markets	Decay	Reasoning
Temps falling, low ahead	×1.0	Full uncertainty — overnight low hasn't occurred
Rising, min ≈ forecast	×0.45	Morning low locked in, matches forecast
Rising, min < forecast	×0.50	Already went below forecast — outcome locked
Midday, stable	×0.65	Morning low is probably set

This means the later scans (11am, 1pm ET) produce tighter, more confident signals than the early morning scan. This is by design — later scans have more data.

Step 3: Bracket Probabilities

For each Kalshi bracket [low, high], the probability is calculated using the normal CDF:

The ±0.5 adjustment accounts for integer temperature reporting. For open-ended brackets (≤X or ≥X), only one CDF bound is used.

Step 4: Edge Detection + Liquidity Filter

A signal is generated when all of the following are met:

• Edge > 15% AND expected value > 3¢ per contract
• Bid/ask spread ≤ 10¢ (rejects thin, illiquid markets)
• Volume ≥ 100 contracts (rejects untradeable markets)
• Price between 3¢ and 97¢ (rejects near-certain outcomes)

The liquidity filter is critical — without it, the model would generate phantom signals on contracts that look mispriced but have no real market to trade against.

Step 5: Position Sizing (Quarter-Kelly)

Where p = model probability, b = payout odds = (1 − price) / price. The 0.25 multiplier is a conservative quarter-Kelly to reduce volatility. Positions are sorted by expected value and allocated until the $100 daily budget is exhausted.

Trading Parameters

Parameter	Value
Daily Budget	$100
Max Per Trade	$20
Min Edge Threshold	15%
Min Expected Value	3¢ per contract
Kelly Fraction	Quarter-Kelly (0.25)
YES Price Floor	≥15¢ (NO stays at ≥3¢)
Min Bid Required	2¢ (ensures exit liquidity)
Max Bid/Ask Spread	10¢
Min Volume	100 contracts
Portfolio Split	70% NO / 30% YES
MAE Suppression	Signals disabled when MAE > 5°F

Understanding MAE (Mean Absolute Error)

MAE measures the average of how far off our forecast is from what actually happened, ignoring direction. If we forecast 52°F and the actual was 55°F, the error is 3°F. If the next day we forecast 60°F and actual was 57°F, the error is also 3°F. The MAE across both days is 3°F.

Why MAE Matters for Bracket Markets

Kalshi brackets are typically 2°F wide (e.g., 52-53°F). The model needs to estimate which bracket the temperature will land in. If our MAE is 2°F, we're usually within one bracket of being right — the probability estimates are trustworthy. If our MAE is 10°F, we're off by 5 brackets on average — the model's probabilities are essentially guesses.

MAE Range	Quality	Action
< 3°F	Strong	Trade with confidence — model is within one bracket
3-5°F	Moderate	Trade selectively — signals have some uncertainty
> 5°F	Too inaccurate	Signals automatically suppressed

Automatic Signal Suppression

The scanner tracks a rolling MAE for each city and market type (HIGH/LOW) based on the last 14 days of settlement data. When a city's MAE exceeds 5°F with at least 5 settlement samples, signal generation for that market is automatically suppressed.

What this means in practice:

• The scanner still scans the suppressed market — it fetches forecasts, observations, and market prices as usual
• The model still runs and calculates probabilities — you can see the bracket analysis on the dashboard
• Settlement data still collects and the calibration continues learning from errors
• But no signals are generated — the market won't appear in the portfolio allocation
• Once the rolling MAE drops below 5°F, signals are automatically re-enabled — no manual intervention needed

This prevents the model from recommending trades in markets where it doesn't yet have enough accuracy to be profitable. As the calibration system learns from more settlements, the MAE naturally decreases and markets re-enable themselves.

Current suppression status is visible on the Roadmap page or via the CLI: node cli.js --calibration. The /api/calibration endpoint also returns the suppressed flag and rollingMAE for each city.

Auto-Calibration

The model improves itself automatically over time through three mechanisms:

• Bias EMA — after each settlement, the forecast error is blended into a running exponential moving average per city. If NYC consistently forecasts 2° too low, the bias adjusts to +2°. Uses α=0.3 so it adapts quickly but doesn't overreact to one bad day. Activates after 3 settlements.
• Rolling Sigma — the standard deviation is recomputed from the last 14 days of actual forecast errors. A city that's been hard to forecast recently automatically gets a wider sigma. Activates after 5 settlements.
• MAE Suppression — the rolling Mean Absolute Error is tracked per city/market. When it exceeds 5°F (with at least 5 samples), signals for that market are automatically suppressed. When accuracy improves below the threshold, signals re-enable. This prevents trading on markets where the model isn't accurate enough.

On first setup, a 90-day historical backtest seeds the calibration using data from NOAA's RCC-ACIS climate database. This gives the model a strong baseline immediately rather than waiting weeks for settlements to accumulate. Run node cli.js --backtest to re-run the backtest at any time.

View the current calibration state with node cli.js --calibration or via the API at /api/calibration.

Data Sources

National Weather Service API

The NWS API is the sole authoritative source for settlement data on Kalshi weather markets. We use five endpoints per city:

• Point forecast — 7-day forecast with daily high/low temperatures
• Hourly forecast — 156-hour temperature forecast, used for stability analysis (24h range, hourly swings)
• Station observations (latest) — current live readings from the exact NWS station Kalshi settles on
• Station observations (today) — all readings today, used to compute observed min/max/trend for the model
• Active alerts — watches, warnings, advisories affecting the city, used in stability scoring

Additionally, the Area Forecast Discussion (AFD) from each NWS office is analyzed for weather pattern keywords.

The NWS API is free, requires no API key, and is rate-limited. We add 300ms delays between requests and implement retry logic with exponential backoff.

NOAA RCC-ACIS (Historical Data)

For the 90-day backtest, we use the Regional Climate Centers' Applied Climate Information System. This provides daily max/min temperature records for each station going back years. No API key required.

Kalshi Public API

Market prices, bracket definitions, and order book data are fetched from Kalshi's public API. No authentication is required for reading market data. Prices are parsed from live bid/ask spreads and last-traded prices.

Polymarket Cross-Reference

In addition to Kalshi, Polymarket offers daily high temperature markets for 36+ cities worldwide, including all 17 US cities we track on Kalshi. The scanner pulls Polymarket prices on every scan and computes a consensus score for each city.

How Consensus Works

For each signal the model generates, we check whether Polymarket's pricing agrees or disagrees:

• BUY NO signal: Our model says Kalshi is overpricing a bracket. If Polymarket prices the same bracket lower than Kalshi, Polymarket agrees the bracket is overpriced → consensus.
• BUY YES signal: Our model says Kalshi is underpricing a bracket. If Polymarket prices it higher than Kalshi, Polymarket agrees it's underpriced → consensus.

The consensus score ranges from -1 (strong disagreement) to +1 (strong consensus):

Score	Badge	Meaning
> 0.5	PM ✓	Polymarket agrees with our signals — higher confidence
0 to 0.5	PM ~	Mixed agreement — some signals confirmed, some not
< 0	PM ✗	Polymarket disagrees — the mispricing may be on our end, not Kalshi's

Why This Matters

A single market can be wrong. But when two independent markets (Kalshi and Polymarket) agree with our model that a bracket is mispriced, the probability of real edge increases significantly. Conversely, if Polymarket's price is close to Kalshi's and far from our model, it suggests our model may be the one that's wrong — two markets independently arrived at a similar price.

Polymarket also offers 11 brackets per city (vs Kalshi's 6), providing finer resolution on market expectations. The scanner matches brackets with overlapping temperature ranges for comparison.

Quick-Reference Trading Guide

Reading the Dashboard Indicators

Indicator	What It Means	Trade with confidence when...
Signal (BUY NO / BUY YES)	Model thinks Kalshi mispriced this bracket	Edge > 20% and EV > 5¢
Bid/Ask	What you'd actually pay. Bid = sell price, Ask = buy price	Spread ≤ 5¢ (tight = liquid market)
PM ✓	Polymarket prices confirm the mispricing	Two independent markets agree against Kalshi
PM ✗	Polymarket disagrees — maybe Kalshi is right	Be cautious, consider skipping
σ [cal]	Sigma computed from real settlement data, not static guesses	Green [cal] = more trustworthy model
Ensemble spread	How much the 6 weather models disagree	Low spread (< 5°) = models agree = more certain forecast
decay	How settled the temperature is based on time of day and observations	Strong decay (×0.4-0.5) = outcome nearly locked in
MAE (via calibration)	How accurate our forecast has been for this city historically	Below 3° = strong accuracy

Strongest Trades

Your highest-confidence trades are when multiple indicators align:

• BUY NO signal with decent edge (> 20%)
• PM ✓ — Polymarket agrees with the signal
• Low ensemble spread — weather models agree on the forecast
• Strong time decay — afternoon local time, temperature near or past forecast
• City with low MAE (< 3°F) — accurate historical track record

When to Skip

• PM ✗ on the signal — Polymarket disagrees, the mispricing may be on our end
• High ensemble spread (> 10°) — weather models can't agree, forecast is uncertain
• City has high MAE (> 4°F) — our forecasts for this city aren't accurate enough
• Wide bid/ask spread (> 8¢) — thin market, hard to enter/exit
• Early morning scan with minimal time decay — too much uncertainty remains

Recommended Workflow

• 6:45 AM PT — Review the landscape. Note which cities have signals. Don't trade yet.
• 8:00 AM PT — Primary trading window. Updated forecast, moderate sigma decay, decent liquidity. Place your best trades here.
• 10:00 AM PT — Rebalance. If existing signals strengthened and PM ✓, add. If they disappeared, the market corrected.
• 12:00-2:00 PM PT — East coast cities are nearly settled. West coast settling. Good for late conviction trades on cities with strong decay.

---

How Settlement Works

Each morning, the scanner fetches the previous day's actual high temperature from NWS observation data for each station. It then:

• Compares the actual high to the forecast to calculate forecast error
• Identifies which bracket the actual temperature landed in (the winning bracket)
• Scores each signal as a WIN or LOSS — a BUY YES wins if the temp lands in that bracket; a BUY NO wins if it doesn't
• Calculates actual P&L (payout minus cost) for each position

Historical performance, win rates, and cumulative P&L are tracked on the History page.

Real Example: Reading the Dashboard

Let's walk through an actual scan to show how to interpret everything the scanner produces. We'll look at Miami HIGH from a recent run, then explain the portfolio allocation.

Step-by-Step: Miami High Temperature

Here's what the scanner showed for Miami's daily high temperature market:

How to Read Each Column

• Bracket — The temperature range. "84-85°" means this contract pays $1 if Miami's official high is 84°F or 85°F. "≥86°" means 86° or above. "≤77°" means 77° or below.
• Model % — Our model's estimate of the probability that the actual high falls in this range. The NWS forecasts 83°F, so the model thinks 82-83° is most likely (42.5%), with 84-85° next (31.3%).
• Bid/Ask — The current Kalshi market prices in cents. The bid is what you can sell at; the ask is what you'd pay to buy. A bid/ask of "-/47¢" means nobody is bidding, but you can buy YES for 47¢. These prices represent what the market thinks the probability is — 47¢ implies a 47% chance.
• Signal — The scanner's recommendation based on comparing our model probability to the market price.

Why BUY NO on 84-85°?

The market is asking 47¢ for a YES contract on 84-85°, implying a 47% chance. But our model estimates only a 31.3% chance. The market is overpricing this bracket — it thinks 84-85° is more likely than our forecast supports.

When a bracket is overpriced, we buy the NO side. A NO contract costs 100¢ − 47¢ = 53¢ (approximately). If the high does NOT land in 84-85° (which our model says is a ~69% chance), we collect $1.00, profiting ~47¢ per contract.

Why BUY YES on 80-81°?

The market is asking only 3¢ for a YES contract on 80-81°, implying just a 3% chance. But our model estimates a 17% chance — more than 5x what the market thinks. This bracket is significantly underpriced.

At 3¢ per contract, you risk 3¢ to potentially win $1.00. With a 17% model probability, the expected value is strongly positive. Even though this bracket is less likely than 82-83°, the price is so cheap relative to the probability that it's a good bet.

Why No Signal on 82-83°?

The model says 82-83° has a 42.5% chance, and the market is asking 39¢. That's close enough — the market and our model roughly agree. The edge (42.5% vs 39%) doesn't clear our 15% minimum threshold, so no signal is generated. Not every bracket needs a trade — we only act when the market is significantly mispriced.

---

Reading the Portfolio Allocation

After scanning all cities, the scanner combines every signal into a single portfolio. Here's how to read it:

Understanding Each Column

• Ticker — The Kalshi contract identifier. You'll search for this in the Kalshi app to place the trade. Example: KXHIGHMIA-26MAR24-B80.5 is Miami's high temp, March 24, bracket centered on 80.5°F.
• Bracket — A human-readable description. "High 84-85" means this is a daily high temperature market for the 84-85°F range. "Low ≥64" means a daily low temperature market asking if the low will be 64° or above.
• Side — YES means you're betting the temperature WILL land in this bracket. NO means you're betting it WON'T.
• Contracts — How many contracts the model recommends buying, based on Quarter-Kelly position sizing. More contracts = higher confidence in the edge.
• Cost — Total dollars you'd spend to buy this position. Each contract costs between 1¢ and 99¢. For example, 120 contracts at 3¢ each = $3.60.
• Exp. Profit — The model's estimate of how much profit this position will generate on average. This is not guaranteed — it's a statistical expectation based on the model probability.

Why So Many Contracts for Miami 80-81° (120) vs 84-85° (14)?

Position size is determined by two things: the edge (how mispriced the contract is) and the price (how cheap each contract is).

• 80-81° YES at 3¢: Very cheap per contract, and the model sees a big edge (17% vs 3%). The Kelly formula says to bet more when the edge is large and the price is low. 120 contracts × 3¢ = only $3.60 at risk.
• 84-85° NO at ~53¢: More expensive per contract. Even though the edge is decent, each contract costs more, so fewer fit within the $20 max-per-trade limit. 14 contracts × 53¢ = $7.56.

What Does "$90.82 Total Cost" Mean?

This is the total amount of money you'd need to place all 10 positions. The scanner targets a $100 daily budget and allocates by expected value — the best opportunities get funded first, and it stops when the budget is used up. You don't have to take every position — you can cherry-pick the ones you're most confident in.

About Expected Profit

The expected profit numbers assume the model's probability estimates are correct. In reality:

• Some trades will win (payout $1.00 per contract) and some will lose (payout $0.00)
• Over many trades, the average profit should approach the expected value if the model is well-calibrated
• Individual days can be profitable or unprofitable — that's normal variance
• The History page tracks actual results vs predictions over time

Tips for Placing Trades on Kalshi

• Search for the ticker or city name in the Kalshi app
• Match the bracket and side (YES or NO) from the scanner
• Use limit orders at or near the bid/ask shown in the scanner, not market orders
• You don't need to trade the exact number of contracts shown — start smaller while you build confidence
• Check the order book depth on Kalshi — if there are only a few contracts available at the ask price, you may not be able to fill the full position
• Later scans (11am, 1pm ET) often have better liquidity and updated forecasts

---

Schedule

Time (PT / ET)	Action
6:45 AM PT / 9:45 AM ET	First scan — early morning forecast, initial market prices
8:00 AM PT / 11:00 AM ET	Second scan — updated forecast, more market liquidity
10:00 AM PT / 1:00 PM ET	Third scan — midday update, strongest observation data
7:00 AM PT / 10:00 AM ET	Settlement check — score yesterday's signals as wins/losses

You can also trigger a scan at any time using the "Run Scan" button on the dashboard. Each run is saved separately so you can compare how prices changed throughout the day.

FAQ