Methodology

Every active series belongs to one of four tiers. A tier is not a quality judgment — it is a statement about when a series carries information relative to closing volume.

• LeadingMoves first; buys weeks of warning. Rates, MBA purchase apps, NAR pending home sales (PHSI), Redfin weekly new listings and price-drop share, NAHB builder confidence.
• CoincidentReal-time confirmation of where the cycle is now. Existing home sales (SAAR), Redfin days-on-market, MBA refi share, and MBA purchase YoY all sit here — they describe current activity, not what is coming next.
• LaggingConfirms a cycle that has already turned. Case-Shiller HPI and ALTA quarterly title premium.
• MacroThe field everything else moves through: 10-year Treasury, 30-year mortgage rate, jobs and prices.

The dashboard groups tiles by tier so a leading signal flashing red cannot be confused with a lagging print rolling over months later.

The headline gauge is computed nightly by compute-health-index.server.ts. For every active series with an index_weight > 0:

1. 01Take the trailing 24 observations and compute the latest z-score: z = (latest − mean) / σ. Series with fewer than 6 observations are skipped.
2. 02If higher_is_better = false (e.g. mortgage rate, days-on-market), flip the sign so "bad direction" is always negative.
3. 03Clamp z to ±3 so one outlier cannot dominate the composite.
4. 04Weight: contribution = z × index_weight. Sum across all contributing series.
5. 05Score: score = clamp(50 + 8 × (Σ contribution / Σ weight), 0, 100). The narrative band assignment in code is: score ≥ 60 reads as expansion, 40 ≤ score < 60 reads as softening, score < 40 reads as contraction. 50 is arithmetically the long-run neutral but falls inside the softening band as currently coded — a known asymmetry we plan to split into a distinct neutral band once we have enough backtest history to justify the cut point.

The "top mover" narrative on the dashboard is the series with the largest absolute contribution, reported in its native unit (basis points, %, $, etc.) — never the internal z-score, which has been a source of misleading "−2.35" lines that looked like raw deltas.

The forecast engine (forecast/compute.server.ts) is deliberately small: no LLM, no neural net, no opaque ensemble. Each horizon is a hand-specified set of lagged drivers, each driver z-scored against its own recent history, weighted, and applied to the target's recent mean as a percentage sensitivity. The target is EXHOSLUSM495S (Existing Home Sales SAAR) — used as a proxy for closing volume because it is the closest national, monthly, SAAR series to actual closings, is revised each release so we can score forecasts against revised actuals, and has the deepest continuous history of the candidate FRED targets (Housing Starts and New Home Sales lead too far upstream of closings to be a clean volume proxy). The 2–4 week horizon projects this monthly target at a sub-monthly tempo by recomputing nightly: each new driver release (MBA weekly, Redfin weekly, PHSI monthly) re-prices the same next monthly print, so the forecast tile updates within the period rather than only at month-end.

The three horizons

Point estimate, interval, direction

1. 01Each driver: read its value at today − lag_weeks, compute z against its last 24 obs, clamp to ±3.
2. 02weighted_z = Σ (z × weight) over contributing drivers only. completeness = Σ |weight used| / Σ |weight defined|.
3. 03Point estimate: point = recent_mean × (1 + weighted_z × SENSITIVITY_PCT), where SENSITIVITY_PCT = 0.05 — a 1σ driver swing moves the projection ~5% from baseline.
4. 04Interval: ±(6% + 3% × high-importance calendar events in window + 5% × (1 − completeness)). Wider when data is sparse, wider again around an FOMC or jobs print.
5. 05Direction: up / down / flat. The flat band is 0.5 × σ of the target's own period-over-period changes — directional calls are reserved for moves larger than typical noise.
6. 06Confidence: clamp((1 − interval × 5) × completeness, 0.1, 0.85). Capped at 85 — we never claim certainty we have not earned. Note the multiplier: interval here is the same fraction from step 04, so under typical conditions (interval ≈ 6–12%) confidence lands roughly 0.40–0.70 at full completeness; if interval climbs past 20% (many in-window events or very sparse data) the formula clamps to the 0.10 floor by design rather than reporting a number we have not earned.

Reliability is measured two ways, and the panel labels which one is currently active.

Walk-forward backtest (backtest.server.ts)

The same lagged-linear model is replayed over historical FRED data with strict no-look-ahead discipline: at each replay date, only observations with obs_date ≤ made_for_date are visible. Split chronologically 70 / 30, grid-tune FLAT_BAND (0.25–1.0) and SENSITIVITY (0.03–0.10) on the in-sample half, measure on out-of-sample. Per horizon we persist: MAE, MAE-ratio vs. naive-baseline, direction-hit-rate, interval-coverage, residual p10 / p90.

Forward scorecard (score.server.ts)

When a forecast's resolves_on date passes, the scorer looks up the actual target print, writes an immutable row into forecast_scores (abs_error, baseline_error, beat_baseline, direction_hit, in_interval), and flips forecasts.resolved = true. Score rows are append-only; we never edit a forecast after the day it was made.

The Reliability card on the dashboard reads the most recent 50 resolved scores per horizon. Once ≥12 live forecasts have resolved and beaten baseline for a given horizon, forward_calibration.active = true takes over from the backtest as the source of truth, and the panel banner switches from "backtest-calibrated" to "forward-measured."

Each observation stores three timestamps: obs_date (the period it describes), as_of (the day the publisher released it), and fetched_at (the day we ingested it). Tiles display the first two so a "live" reading from three weeks ago is not confused with a fresh print. A tile is marked stale when its latest observation is older than its frequency would allow (weekly: 16 days, monthly: 75 days). A tile shows no data when no observations exist yet.

National series — inputs to the Health Index

These are the 40 national series the Health Index actually composes. Tier here is the tier the production code assigns — if a series appears in a tier that surprises you, the catalog reflects the code's classification, not the prose above.

Metro series — local view only (not in the national Health Index)

The 130 metro-level rows below power the Local page only. They do not flow into the national Health Index or any forecast horizon. They are listed for transparency about what we ingest, not as components of the national gauge.