Data Quality & Dashboard Fixes

Problem: All five market data pages were statically cached at build time. New Supabase data (permits through Dec 2025, HPI through Q3 2025) was invisible to users.

Fix: Added export const revalidate = 3600 to each page, enabling Next.js Incremental Static Regeneration. Pages now re-fetch data at most hourly.

Pages affected:

• /market/[cbsa] — Market Overview
• /market/[cbsa]/permits — Permits
• /market/[cbsa]/employment — Employment
• /market/[cbsa]/demographics — Demographics
• /market/[cbsa]/prices — Prices

Verification: Build output shows all market pages as ƒ (dynamic/ISR) instead of ○ (static).

Problem: The finished_vacant_modeled metric subtracted HMDA purchase originations (~106K/year, all housing types) from new-construction completions (~18K/year, single-family only). The result was always deeply negative — off by approximately 5×.

Fix: Replaced the Finished Vacant KPI card with a “Closings (HMDA)” card showing the raw quarterly purchase originations. The underlying finished_vacant_modeled column remains in the database for future work when a better closings proxy (new-construction-only HMDA filter or recorder deed transfers) becomes available.

Problem: When BPS data had a month count not divisible by 3 (e.g., 13 months), the pipeline model produced a final “quarter” with only 1–2 months of data. This partial quarter appeared artificially low on the dashboard.

Fix: Added a check at the end of run_quarterly():

if n_months % 3 != 0 and quarters:
    quarters.pop()

Test: New test test_incomplete_quarter_dropped verifies that 13 months produce exactly 4 full quarters, not 5.

Result: Pipeline now outputs 8 clean quarters for Houston (Jan 2024 – Dec 2025).

Problem: The getPipelineQuarterly query fetched from mart_pipeline_quarterly without filtering by model_version. If multiple model runs existed, the dashboard would show duplicate quarters.

Fix: Added .eq("model_version", "v1") to the Supabase query.

Problem: fact_bls_employment had no foreign key to dim_cbsa, and common dashboard query patterns lacked indexes.

Fix: Migration 013 adds:

• FK constraint: fact_bls_employment.cbsa_code → dim_cbsa.cbsa_code
• Index: (cbsa_code, ref_month) on BLS table
• Index: (cbsa_code, index_type, ref_period) on HPI table

Problem: The pipeline model operates on units_1u (single-family only), but KPI cards said “Quarterly Permits,” “Starts,” etc. Meanwhile the “Monthly Permits” card showed units_total (all types). Users couldn’t tell which metric was which.

All numeric values also gained .toLocaleString() formatting for consistent thousand-separators.

Problem: The permits bar chart showed “Single Family” and “Total Units” as separate side-by-side bars. Since Total includes SF, this visually double-counted single-family permits.

Fix: Replaced with stacked bars showing Single Family (bottom, dark blue) and Multifamily (top, light blue), where Multifamily is computed as units_total − units_1u. The stack total equals the real total.

Problem: Census PEP 2020 data covers only April–June (3 months instead of 12). In the “Components of Change” chart, births appeared to collapse from ~90K to ~22K — a misleading artifact of the short coverage period.

Fix: Filtered 2020 from the components chart (ref_year > 2020). The population trend chart retains 2020 since the estimate itself (7.17M) is correct.

Problem: The demographics page hardcoded a + prefix on YoY population change and net migration. If a metro shrank, the display would show +-1,234.

Fix: Conditionally prepends + only when the value is non-negative.

Problem: The mortgage rate chart loaded all weekly FRED data back to 1971 (~2,800 data points). The chart was dense and slow, and historical rates from the 1980s weren’t relevant to current market analysis.

Fix: Added an optional startDate parameter to getFredSeries() and filtered mortgage data to 2020–present.

Problem: Unemployment rate (3–6%) and mortgage rate (5–8%) charts forced the Y-axis to start at 0. This compressed meaningful variation into the top sliver of the chart area.

Fix: Changed Y-axis domain from [0, "auto"] to ["dataMin - 1", "auto"]. The axis now starts just below the lowest data point, giving better visual resolution of rate movements.