Methodology

How BlueSkyFI grades the rate signal track record

Money Weather explains financial conditions as changing weather, with uncertainty attached. Our mortgage track record lets you inspect the original call and the later rate observation. The current era begins on 2026-05-19; results through 2026-05-18 remain in a separate archive. Changing a model does not erase misses or rewrite earlier grades.

The source and the clock

We check the daily 30-year conforming mortgage-rate series OBMMIC30YF on FRED. Treasury yields and other market signals help explain conditions; they do not replace the mortgage observation being scored. This benchmark is not a personal lender quote.

  • Each call retains its published date, original recommendation, confidence, explanation, and call-time mortgage-rate baseline. The source observation date can be earlier than the call date.
  • The target is five business days after the forecast call date in America/New_York, not five days after the baseline observation. The calendar counts Monday through Friday; missing holiday observations are handled by the source window.
  • New grades and the strict forecast comparison use the first valid daily FRED OBMMIC30YF observation from the target date through one additional business day. The actual observation date is required; a later reading cannot substitute for the intended window. Earlier grades remain preserved, with exclusions disclosed in the strict comparison.
  • The backfill process waits one extra business day for source publication. Pending means the outcome window has not matured; awaiting backfill means an eligible outcome still needs to be retrieved. Neither is a miss or a win.
  • Malformed, fallback, placeholder, stale, unavailable, or contaminated rate observations are excluded with a reason. A missing observation is never filled with a guessed rate to improve the score.

Two scores with different meanings

Historical action grades keep their original rule: lock is correct at a rise of at least 0.03 percentage points, float at a fall of at least 0.03, and hold within the inclusive ±0.03 band. Exactly at either boundary, hold and the matching directional action can both count as correct. Directional accuracy includes lock and float; all-call accuracy also includes hold.

The separate forward forecast evaluation assigns exactly one outcome: up above +0.03 percentage points, down below −0.03, and flat between those limits, including both boundaries. We round the rate change to three decimal places before applying the rule. A move from 6.40% to 6.43% is flat; a move to 6.44% is up. We compare this outcome with the forecast recorded at the time and report ambiguous forecasts as exclusions.

Forecast direction and household action answer different questions. Rates may be more likely to rise while the practical recommendation remains hold because evidence is weak or a household needs more information. We evaluate the forecast separately rather than treating every hold as a prediction of flat rates. These forecast scores do not replace historical action grades.

What Gemini contributes

Gemini leads the rate outlook when usable evidence is available, assesses the supplied evidence, weighs opposing explanations, and translates the result into household language. Its explanation should identify what changed, the dated source, how it could affect mortgage pricing, who it affects, and what new evidence would change the call. Linked reporting is context; Gemini cannot create a missing numeric observation. Calculations and grading remain deterministic, and unavailable or invalid evidence can require an abstention.

The rules-only baseline, Gemini assessment, and published decision remain separate audit records. This helps us check whether the added analysis improves decisions. Model confidence is an estimate, not measured accuracy, and a persuasive explanation is not proof of predictive skill. A change in approach must be tested against new outcomes before we can claim it improves accuracy.

Calibration uses comparable, version-specific forward evidence. Unmatched historical groups cannot establish improvement or justify transferring an old accuracy estimate to a new approach. We disclose performance and data requirements without publishing proprietary prompts, settings, or decision weights.

Coverage, comparisons, and uncertainty

Insufficient evidence is an abstention, not a flat forecast or a pending outcome. We report missing and abstaining forecasts alongside accuracy. Coverage is the share of eligible outcomes with an evaluable forecast; abstentions and missing forecasts remain in that denominator. A high score on a small fraction of days does not describe performance on every day.

We compare rules, Gemini, and published decisions on matching dates, with simple always-up, always-flat, and always-down benchmarks in the forward view. We track forecasting versions separately so a change in approach cannot silently inherit an earlier score. Dated records help keep comparisons fair, but timestamps alone cannot prove that every source was available when a call was made.

Daily five-day forecasts share much of the same market window. We also check periods that do not overlap, so several calls riding the same rate move are not mistaken for many independent successes. Even these periods can share market conditions. Statistical intervals describe uncertainty around a score; they are not promises about the next forecast. A promising change needs testing on new outcomes, not just the history that suggested it.

What these results can tell you

95% accuracy is a goal, not an established result. We do not claim a success rate for an updated approach before measuring it on new, matured outcomes. We look for durable results across different conditions, with coverage, simple benchmarks, and uncertainty visible alongside the headline score. Mortgage results cannot establish accuracy for inflation, savings, investing, or other Money Weather signals; each needs its own outcome definition and forward evaluation. The purpose is better-informed household choices, not a promise of perfect market timing.

BlueSky FI | Rate Signal Methodology