About Datum Signals
A public, behavioral mood index for every US state, refreshed daily. Built to answer a different question than polling: not what voters say, but what their behavior reveals.
Why this exists
Polling tells you what people will admit to a stranger on the phone. Behavioral data tells you what they actually do. Voter registration filings, unemployment claims, vehicle miles traveled, gas prices, search interest in safety topics, suicide rates: these are recorded by neutral bureaucracies and reported on cadences nobody can game. They form the foundation of a civic mood index that is descriptive first, predictive second, and transparent always.
The forecast industry as it stands is split between two flawed extremes. Expert-judgment shops (Cook, Sabato, Inside Elections) deliver narrative ratings that can move on a single staffer's hunch. Quant shops (538, Decision Desk, prediction markets) lean on polls and fundamentals but rarely show their math in detail. Both have value. Neither shows you how civic mood is actually moving on the ground between elections.
Datum Signals occupies the gap. The index runs all year, every year. Methodology is published in full. Brier scores on every back-tested cycle since 2014 are public. When the model gets a race wrong, the correction goes in /corrections.
Methodology integrity
The composite mood score for each state is built from 14 behavioral dimensions covering mobility, security, economy, campaign spend, civic trust, domestic stability, isolation, risk appetite, escape behavior, information consumption, candidate-quality signals, donor energy, and polling pulse. Each dimension is normalized against the other fifty states, then combined with calibrated weights, on top of which the forecast applies a long-run midterm correction. Accuracy is reported against the full record back to 2014, scored out-of-sample cycle by cycle.
The forecast layer is labeled experimental and carries its current Brier score everywhere it appears. That score is currently 0.119, measured out-of-sample across every cycle back to 2014 rather than on a single favorable split, versus 0.250 for a coin-flip baseline. It has improved from 0.247 at first release through a sequence of documented changes, including the ones that made it worse.
The index is recalibrated quarterly. Any change to a feature definition, a coefficient, or a data source is logged before it goes live, and the previous model version remains queryable so independent reviewers can reproduce older calls.
Who built this
Datum Signals is built and operated by David B. Wheeler (Big D), founder of VoteROI.com LLC. He's run for office four times in North Carolina, hosts the MUCK YOU podcast, and has visited 75+ countries across all seven continents. Datum Signals is designed to make civic mood data available to journalists, researchers, campaigns, and anyone else who'd rather work from primary behavioral data than aggregated punditry.
The platform is headquartered in Spruce Pine, North Carolina. The product voice is deliberate: the National Weather Service for civic emotions. Forecasts, advisories, watches, conditions. Honest about uncertainty, allergic to spin.
How we differentiate
- vs Cook Political Report
- They publish expert ratings; we publish a model with a Brier score. Both are useful. Theirs reflects judgment, ours reflects measurement.
- vs FiveThirtyEight
- 538 combines polls plus fundamentals into a probabilistic forecast. We use no polls. The two are intentionally complementary: when polls and behavior diverge, that's a signal both signals are worth knowing.
- vs Kalshi KPOW / Polymarket
- Prediction markets price expert and amateur belief. We measure what's happening on the ground. Markets are good at near-term aggregation; behavioral data is better for the months between elections when polls go dark.
- vs Decision Desk HQ
- DDHQ is a fundamentals plus polling shop with a fast call desk. We don't call races; we report mood. Different products for different needs.
What we will and will not do
We publish what the model says. We don't shade calls to match a preferred narrative. We don't accept consulting work that requires us to bend the methodology. We don't run AI-generated commentary; every line of editorial copy is written by a human, with Claude assisting on drafts and the final voice human. When we get something wrong, we say so in /corrections, with the date, the version, and the fix.