A FRESHMAN GUIDE TO PREDICTION SCIENCE

When does data become
a signal?

Follow one hourly crypto observation as it travels through prices, percentages, momentum, uncertainty, and risk checks. No magic. No crystal ball. Just measurements, rules, and honest testing.

15 MIN READALGEBRA + STATISTICSINTERACTIVE LAB

01 / THE BIG IDEA

A prediction is a claim that can be tested.

Suppose a Kalshi market asks whether ETH will finish above a particular price, called the strike. A YES contract pays $1 if the event happens and $0 if it does not. A NO contract represents the opposite outcome.

Our hourly favorite model does not claim to know the future. It asks a narrower question: Is one side already ahead, moving in the same direction, sufficiently far from the strike, and trading in a calm, liquid market?

02 / THE PIPELINE

One signal. Nine gates.

Every gate must pass. In logic, this is an AND rule: one failed condition is enough to produce “no signal.”

Clock

10–58 minutes remain

Leader

YES or NO is ahead now

Distance

Spot buffer is sufficient

Motion

5-minute trend agrees

Freshness

Observation ≤ 120 sec old

Price

Contract costs $0.75–$0.90

Spread

Ask minus bid ≤ $0.02

Volatility

Movement is not too noisy

Liquidity

At least 1 contract exists

hourly_trigger.py · simplified

eligible = (
    10 <= minutes_left <= 58
    and ask_depth >= 1
    and currently_winning
    and momentum_matches_side
    and latest_price_is_fresh
    and 0.75 <= contract_price <= 0.90
    and spread <= 0.02
    and spot_buffer >= asset_minimum
    and volatility <= asset_maximum
)

03 / THE MATHEMATICS

Four small ideas do the heavy lifting.

Pick the current leader

YES if S > K
NO if S < K

S is the Coinbase spot price. K is the Kalshi strike. The model follows the side currently ahead; it does not bet on a comeback.

Measure the lead

buffer = |S − K|S × 100%

This turns a dollar distance into a percentage, allowing us to compare differently priced assets fairly.

Confirm momentum

r₅ = ln(S_nowS_{5m ago})

YES needs a positive return. NO needs a negative return. A logarithm makes upward and downward proportional moves easier to compare.

Measure uncertainty

σ = √Σ(r_i − r̄)²n − 1

Volatility is the sample standard deviation of one-minute log returns. Bigger σ means more scattered, less predictable movement.

ASSET-SPECIFIC GUARDRAILS

More movement requires more room.

DOGE is permitted more minute-to-minute volatility, but it must also be farther from its strike. These are model settings—not laws of nature—and must be evaluated with new data.

Asset	Minimum buffer	Maximum volatility
ETH	0.20%	0.10% / minute
XRP	0.25%	0.12% / minute
DOGE	0.30%	0.20% / minute

04 / WORKED EXAMPLE

Walk an ETH signal through every gate.

CANDIDATE PASSES

SPOT NOW$2,510

STRIKE$2,504

SPOT 5M AGO$2,500

YES ASK / BID$0.79 / $0.78

Spot buffer

|2510 − 2504|

2510 = 0.00239 = 0.239%

ETH needs at least 0.20%. PASS

Five-minute return

ln(2510 ÷ 2500)= 0.00399 = +0.399%

Positive motion agrees with YES. PASS

Market spread

$0.79 − $0.78= $0.01

The maximum is $0.02. PASS

✓ 35 MIN LEFT✓ YES AHEAD✓ FRESH DATA✓ 0.08% VOLATILITY✓ 4 AVAILABLE✓ $0.79 PRICE

05 / INTERACTIVE LAB

Change the evidence.
Watch the conclusion change.

These controls use sample observations only. Move one value outside its allowed range and see why the candidate is rejected.

SIMULATION ONLY · NO LIVE DATA · NO ORDERS

06 / THINK LIKE A SCIENTIST

A signal is the beginning of an experiment—not the end.

Form a hypothesis

“When all nine conditions pass, the favorite may win often enough to overcome its entry cost and fees.”

Record every trial

Keep wins, losses, skipped signals, timestamps, quotes, fees, and actual fills. Selective memory is not evidence.

Test unseen data

A rule tuned to old results can memorize noise. Future, out-of-sample observations provide a more honest test.

Quantify uncertainty

A small sample can look extraordinary by chance. Confidence intervals help show how uncertain the estimated win rate remains.

THE MOST IMPORTANT EQUATION

Winning often is not the same as making money.

break-even win rate ≈total debit ÷ $1 payout

If a contract costs $0.78 and the fee is $0.01, total debit is $0.79. A win earns approximately $0.21; a loss costs $0.79. The strategy needs roughly a 79% win rate on comparable entries just to break even. This is why price, fees, and calibration matter as much as accuracy.

Field notes

Spot: The asset’s current observed market price.
Strike: The threshold named in the prediction contract.
Spread: The difference between the lowest ask and highest bid.
Liquidity: How readily contracts can be bought or sold.
Volatility: A measurement of how dispersed price changes are.
Calibration: Whether events predicted at a given probability occur about that often.

YOUR NEXT QUESTION

“What evidence would change my mind?”

That single sentence is the bridge between guessing and science.

CONTINUE TO LESSON 02: BUILD THE LAB →