How NorthStar works

Where the numbers come from, how they are transformed, what the models do, and where their limits are. Data last refreshed Sep 25, 2026, 10:49 a.m. ET.

Data sources

Pipeline and refresh

A GitHub Actions job runs at 10:00 ET every weekday, after the 8:30 ET statistical releases. It ingests all sources, builds a DuckDB warehouse in three layers (raw, clean, marts), exports JSON and Parquet files, and commits them. The commit triggers a new Vercel deployment, so every page is pre-rendered with the latest data. The run stops before deploying if any source fails or if a data-quality check fails (missing series, duplicate rows, data older than 120 days), so the site keeps showing the last good data instead of a broken update.

Transformations

• Frequencies. Daily and weekly series are averaged to monthly. Policy and prime rates use the rate in force at month end. Dates mark the start of each month or quarter.
• Month to date. Series built from daily data include the current, incomplete month; the dashboard labels these readings "Month to date".
• Growth rates. Year-over-year change compares with the same period a year earlier. Quarterly GDP growth is annualized: ((this quarter ÷ last quarter)⁴ − 1) × 100.
• Spliced exchange rate. The Bank of Canada CAD/USD series starts in 2017; earlier months use FRED's USD/CAD rate, inverted.
• Policy rate before April 2009. Filled from the Bank of Canada's published rate decisions, entered by hand.
• Recessions. U.S. dates follow the NBER (FRED series USREC). Canadian dates follow the C.D. Howe Institute Business Cycle Council and are entered by hand.
• Relative strength. A sector ETF's return minus its market benchmark's return over the same window (S&P/TSX 60 via XIU for Canada, S&P 500 via SPY for the U.S.), in percentage points.
• Derived indicators. Real policy rates subtract core inflation (CPI-trim; core PCE) from the policy rate. Real wage growth subtracts headline CPI inflation from wage growth. GDP per working-age person divides real GDP by the quarterly average of the LFS population aged 15+ (Canada) or the civilian population aged 16+ (U.S.). The Sahm rule is the 3-month average unemployment rate minus its lowest 3-month average over the previous 12 months. Canada's breakeven inflation is the long-term benchmark yield minus the Real Return Bond yield.
• 3-month annualized rates. ((this month ÷ three months earlier)⁴ − 1) × 100.
• CAD–oil correlation. Rolling correlation of monthly percentage changes, computed only over complete 12- or 36-month windows.

Models

• Forecasts (Amazon Chronos-2). A pretrained time-series foundation model used zero-shot, without training on this data. It produces 12-month paths at the 10th, 50th and 90th percentiles. Each series is back-tested on its last 12 months against a no-change forecast, and the result is shown next to the chart.
• Rate-shock effects (DoubleML). A partially linear model estimates how a Bank of Canada rate change relates to same-month sector returns, after removing what U.S. inflation and rates, oil, the Canadian dollar, unemployment and recent returns explain. Random forests handle the controls, with 5-fold cross-fitting. Effects are reported per 0.25 percentage point with 95% confidence intervals.
• Macro radar. Not a model: each pillar is today's reading as a percentile of its own history since 2006.

Model outputs are refreshed by re-running the Colab notebook, not daily. The date of the last run appears on the Forecast & scenarios page.

Ask the data

The assistant sends your question and a data dictionary to Claude, which writes one SQL query. The query runs in an in-memory DuckDB database holding only the dashboard's tables, with file, network and settings access switched off. Only single read-only SELECT statements are accepted, results are capped at 1,000 rows, and a second model call summarizes the rows it returned. The SQL is always shown so answers can be checked.

Limitations

• Canadian retail sales in this table start in 2017. The Staples and Utilities ETFs in Canada start in 2012.
• Canadian ETFs are priced in Canadian dollars and U.S. ETFs in U.S. dollars, so cross-border return gaps include currency moves.
• Retail sales are nominal, so inflation lifts their growth rates.
• The causal estimates rest on about 20 years of monthly data and relatively few policy moves, and markets often price moves before they happen. Most effects are not statistically significant; treat them as exploratory.
• Chronos-2 forecasts carry no economic structure. They extrapolate patterns and can miss turning points.
• Statistical agencies revise past data. The dashboard always shows the latest vintage, not what was known at the time.
• Nothing on this site is investment advice.

Series catalog

Every series on the dashboard, generated from the live data.