Forecast core: test results

6 November 2026. The combined model did not improve the forecast. The live service keeps corrected guidance, with trend fallback.

12 historical events at six B2B companies

Calendar Q1 and Q2 2025 at Palantir, NICE, Pegasystems, PROS, Teradata and Manhattan Associates. The candidate combines trend, damped trend, seasonal growth and eligible guidance. Weights use only earlier errors. Later outcomes do not set those weights.

Method	Mean scaled error ↓	Mean error, USD million ↓
Simple trend	0.786	25.93
Corrected guidance with trend fallback	0.630	20.21
Combined model	0.735	23.31

The candidate had 16.6% higher mean scaled error than the current method. It won five cases and lost seven. Guidance was available for two target quarters; the reference method used trend in the other ten. This is a test of the complete fallback method, not twelve guidance forecasts.

A data error invalidated the first score

The first run mixed thousands and millions in Pegasystems revenue. We retained and invalidated that run. We then read the units from the financial statement header and checked available SEC figures. The model, weights, events and thresholds stayed fixed. These corrected results are a mechanical repair after outcomes were inspected. They are not a second independent test.

All ten available SEC outcome comparisons matched. Two NICE outcomes used the original release without a structured SEC comparison.

Range results

The candidate range contained ten of twelve outcomes. Its mean width was 21.8% of the point forecast. Twelve observations across six companies are too few to establish general probability accuracy. This range is not in the live product.

What changed in the software

Stop a forecast if consecutive revenue values differ by about 1,000 times. This catches some unit errors; it does not verify every input.
Show the number of source documents, accepted guidance ranges and earlier forecast tests.
Show earlier forecast errors and the limits of each result.
Keep the current forecast values. Do not deploy the candidate that lost this test.

A smaller guidance-policy candidate was stopped at the input check: it could not affect enough companies. It was not scored.

Cost and scope

Paid project API calls: 0. API cost: $0. Engineering, chat and hosting costs are excluded. These tests measure total revenue under a published business plan. They do not measure customer behavior or the effect caused by a proposed decision.

Method references: forecast combinations and time series cross-validation. These references support the test design, not a claim that this product has state-of-the-art accuracy.