Our Methodology

Market Signals is built on one principle: a forecast is worthless without a way to measure it.

How we form probabilities

1. Evidence-first modeling. Every probability starts with raw data - on-chain metrics, economic indicators, sentiment feeds, historical baselines - not hunches or headlines. 2. Explicit priors. We state our baseline assumptions and what evidence shifts them. If a forecast changes, you can see why. 3. Granular time horizons. Every prediction carries a resolution window (e.g., "by March 31, 2025", "within 24 hours of announcement"). Vague deadlines enable vague accountability.

How we score forecasts

We use the Brier score as our primary scoring rule, complemented by log-loss for probabilistic grids.

• Brier score ranges from 0 (perfect) to 1 (completely wrong). It penalizes overconfident errors more harshly than naive guesses. A probability of 0.9 assigned to an event that does not occur scores 0.81 - much worse than saying "I do not know" and assigning 0.5.
• Log-loss applies when we provide full probability distributions (not just binary outcomes). It rewards honest uncertainty: saying "60% yes, 40% no" is properly rewarded if the event occurs; it is not punished as harshly for being uncertain but correct.

Every forecast receives a score at resolution. We publish both the prediction and the result so readers can verify our scoring independently.

Our commitment

We will never delete a miss. Resolved forecasts - correct or incorrect - are permanent record. Unresolved forecasts are marked with their original timestamp and time horizon visible. If we get something wrong, you will see it alongside what went right. That is the minimum honest forecasting requires. We publish every resolved forecast.