How the Seatbelt is built

Every day, automatically.

1. We fetch

~50 free public series (Treasury yields, credit spreads, the dollar, volatility, inflation, housing, options-market sentiment) from FRED, Cboe and free market data. Nothing is typed by hand.

2. We normalize

Each indicator is scored against its own last year: how far from normal (level z-score), which way it is moving (trend z-score), and whether that helps or hurts risk (a fixed direction). The blend is squashed onto a smooth −2…+2 scale, so small daily shifts move the meter instead of being swallowed by a step function.

3. We weight by confidence

Live data counts full. Delayed, proxy or stale data counts less. Dead feeds count zero, and we tell you.

4. We assemble the Send Meter

A macro core, minus fragility, session and options penalties, plus a confirmation bonus, minus a data-confidence penalty. Scored 0 to 100.

5. We translate

The number becomes a color verdict, a main risk, a named trap and one rule, in trader English.

6. We log it

Every verdict goes to a public ledger and gets an honest end-of-day review.

Validation & sources

Does the signal work?

A risk preflight should predict downside risk, not price direction. We backtested the live scoring math (model v1.3.0: winsorized level z + trend z squashed by tanh, plus a 12-lookback trend-regime vote and dip-in-uptrend conditioning) over 9,122 trading days of FRED history: 1990 to 2026, four bear markets. Full sample (NASDAQ, forward 20 days):

Risk-off days

25.3%

forward 20-day volatility

15%

odds of a >10% drop in a month

−4.6%

average drawdown

Risk-on days

15.7%

forward 20-day volatility

odds of a >10% drop in a month

−2.4%

average drawdown

The middle tercile lands in between (17.7% vol, 6% drawdown odds), so the separation is monotonic across the whole 36 years, including the 2000, 2008, 2020 and 2022 bears. Earlier model versions only separated in the recent cycle; v1.3.0's trend-regime vote (12 momentum lookbacks, an approach validated in live trading research) is what fixed the full-sample robustness. In the recent cycle the spread is wider still: 28.4% vs 15.1% vol, 18% vs 2% drawdown odds. Backtest uses a 4-input proxy with full 36-year history; the live engine runs ~50 inputs including credit spreads, the dollar, and financial-stress indices.

Proxies, labeled honestly.

Where a pro index isn’t free (MOVE, ISM, CRB) we build a transparent proxy and say so. What we never do: hand-edit a score, paste a number, or override the engine. The Send Meter formula: 50 + 12·MacroCore − Fragility − Session − Options + Confirmation − DataConfidence.