ARIMA Forecast Model
Our ARIMA(1,1,1) model produces proprietary estimates for major economic indicators before each release. Here's exactly how accurate it is — including where it falls short.
Where our model is strong
These indicators are well-suited to ARIMA modelling — they follow stable trends with low volatility. Our backtested accuracy on these is genuinely competitive with paid consensus estimates.
CPI
0.17%
MAPE · 36-month window
Index level series. Worst case error: 0.72%. Trained on BLS historical data.
PPI
0.34%
MAPE · 36-month window
Index level series. Worst case error: 0.87%. Trained on BLS historical data.
Jobless Claims
0.20%
MAPE · 12-week window
Weekly initial claims. Worst case error: 0.50%. Sourced from FRED (ICSA series).
Live track record — independently verifiable
Every estimate below was generated by our ARIMA model before the official release date. Each row links to a timestamped GitHub commit proving the estimate existed before the print — no post-hoc fitting. Historical rows use walk-forward backtest methodology; live rows are generated weekly every Monday morning.
Loading estimates...
Note: Rows showing "generated 2026-07-07" were produced via walk-forward backtest on that date — the model had no access to future data, but the timestamp does not reflect a live pre-release generation. Live estimates generated from July 10 2026 onwards carry genuine pre-release GitHub commit timestamps.
Estimates committed to github.com/filingapi/filingapi-estimates every Monday at 07:15 UTC. Click any verify link to inspect the raw commit timestamp on GitHub.
Full model performance
Walk-forward out-of-sample backtest across all indicators we cover. MAPE is calculated on the level series, not month-on-month change.
| Indicator | Window | MAPE | Worst case | Notes | Status |
|---|---|---|---|---|---|
| CPI | 36 mo | 0.17% | 0.72% | Index level, not MoM change | Strong |
| PPI | 36 mo | 0.34% | 0.87% | Index level, not MoM change | Strong |
| Jobless Claims | 12 wk | 0.20% | 0.50% | Weekly initial claims level | Strong |
| NFP | 36 mo | 0.07% | 0.15% | Total payroll level (~158,000k), not monthly change. Direction accuracy is weak — an industry-wide problem. | Level only |
| GDP | — | — | — | Quarterly release · next due Jul 30. Limited historical data for backtesting — indicative only. | Indicative |
MAPE comparison
Visual comparison of model accuracy across the indicators we're confident in.
Model MAPE by indicator
Lower is better · walk-forward out-of-sample backtest
Methodology
All estimates are generated by our ARIMA(1,1,1) model trained on government data sources. These are not Wall Street surveyed consensus estimates — they are our own model output, clearly labelled as such in every API response via the estimate_model field.
Model type
ARIMA(1,1,1) — AutoRegressive Integrated Moving Average
Backtest methodology
Walk-forward out-of-sample — no lookahead bias
Data sources
BLS API (CPI, PPI, NFP), BEA API (GDP), FRED (Jobless Claims, FEDFUNDS)
Update frequency
Weekly refresh · actuals ingested automatically on release day
Where we're improving
We've deliberately excluded indicators where our model isn't strong enough to be useful. These are on our roadmap for more sophisticated modelling approaches.
NFP monthly change direction
Predicting the direction of month-on-month payroll change is notoriously difficult — even Wall Street consensus misses frequently. We're exploring ensemble approaches and alternative leading indicators to improve directional accuracy beyond coin-flip.
Unemployment rate
Our current ARIMA model for unemployment produces weak directional accuracy. We're working on a more accurate approach before surfacing estimates publicly.
ISM / PMI
Currently excluded due to data licensing restrictions from the Institute for Supply Management. We're evaluating alternative sources.
International macro indicators
ECB, BoE, and other central bank release calendars are not yet covered. On the roadmap as we expand beyond US data.
* MAPE (Mean Absolute Percentage Error) is calculated on the level series, not month-on-month change, unless otherwise stated. NFP MAPE of 0.07% refers to the total non-farm payroll level (~158,000k workers), not the monthly change figure reported in headlines. All backtests use walk-forward out-of-sample methodology with no lookahead bias. Estimates are generated by FilingAPI's proprietary ARIMA model and are not Wall Street consensus surveys. Past model accuracy is not a guarantee of future performance.