Model performance

Did the trend beat “same as last year”?

Promotion rule: significantly lower MASE than naive_last_v1 in at least 2 of 3 evaluation periods (1-year horizon). Result: the straight-line trend was significantly better in 0 of 3 periods, so the reference model is Same as last year. Neither model is validated for decisions; both stay labelled experimental in reports.

Accuracy against first-reported revenue

Each forecast is compared with the revenue the company first reported for the target year. Errors are scaled by how much the company's revenue typically changes from one year to the next (MASE), so large and small companies count equally. A MASE of 1.0 means the forecast missed by about one typical year-to-year change.

Against revenue as later restated, MASE is same as last year 2.26 (1 year) and 4.85 (2 years); straight-line trend 2.84 (1 year) and 4.89 (2 years). 21 of 158 scored target years (13.3%) were later restated by more than 0.05%.

By evaluation period

Forecasts grouped by the date they were made. 1-year horizon unless noted.

By sector (1 year ahead)

Sectors use the SEC industry classification (SIC division) of each company today.

Error ranges

FinSavoir shows an error range next to a forecast only if a range calibrated on earlier years held up on later, separate years: 80% ranges are calibrated on outcomes that were public before Jan 1, 2020 and tested on forecasts made after it. The test must include enough forecasts and land within ±5 points of 80% coverage.

No range passed, so reports show no error range for either baseline.

By company (1 year ahead)

Average absolute error in each company's reporting currency, and the straight-line trend's error relative to “same as last year” on the same forecasts (below 1.00 means the trend was more accurate for that company).

Prediction ledger: forecasts FinSavoir has issued

Everything above is a simulation of the past. The ledger is different: each time new filings arrive, FinSavoir records its baseline forecasts with a timestamp, the exact inputs and the model version. Records can never be edited or deleted. When the company reports the target year, the actual value is added and the forecast is scored.

None of the issued forecasts can be scored yet: the earliest target fiscal year ends around Sep 30, 2026, and FinSavoir scores a forecast only after the company files its annual report. 8 earlier forecasts have been superseded by newer ones for the same period and model; they are kept unchanged and will still be scored.

Limitations

• Survivorship bias. The companies are today's S&P 500 and Nasdaq-100 members, not the index members at each past date. Companies that shrank, were acquired or failed are missing, so results likely look better than a real-time universe would have.
• Short history. Machine-readable SEC data starts in 2009–2011, so early forecasts use only 3–5 years of history and each company contributes at most about 16 annual forecasts.
• Excluded companies. 0 of 7 companies have no usable annual revenue series in SEC data (new registrants, or revenue tagged in a way FinSavoir does not map) and are not scored. Forecasts whose target year is missing or whose fiscal calendar changed are excluded, not guessed.
• Pooled results. Scaled errors are averaged across very different companies; a model that is better on average can still be much worse for a particular company.
• These are mechanical extrapolations of reported revenue. They are not predictions of company performance or investment advice.

Backtest run 8209b736 · protocol revenue_fy_walkforward_v1 · normalization v8 · completed Oct 6, 2026. How the data is prepared: methodology.