Methodology: Financial Forecast Generation
This document describes what Sigmodx records for forecast generation agents, how reliability is computed against post-period actuals, and what the attestation verification string proves.
What Sigmodx records
For each forecast, the agent submits a decision (submit, revise, or reject) with an input hash, rationale, confidence, forecast type, forecast period, forecast value, confidence interval, model version, and a hashed list of input data sources. After the period closes, the customer records the actual value, which Sigmodx uses to compute the forecast error. Raw source data and underlying records remain in the customer environment.
Regulatory context
Forecasts and estimates that feed the financial statements fall under SOX internal control requirements and PCAOB scrutiny of management estimates. The append-only decision log and deterministic attestation hash give auditors a tamper-evident record of which AI agent produced a forecast, what it was based on, how confident it was, and how it performed against the real outcome.
Input hashing
Input hashes should include stable identifiers and features such as forecast period, model version, and references to the input data sources. Do not include raw customer records, proprietary model weights, or non-public financial detail in the hashed payload.
Materiality and escalation
Forecasts flagged as material to the financial statements are routed for immediate review regardless of queue position. Each organization configures its own materiality threshold, and the materiality flag rate is tracked as a reliability signal.
Reliability signals
Five signals are computed per period and inserted append-only:
- Forecast accuracy: the fraction of forecasts within 10% of the recorded actual, computed once post-period actuals are available.
- Calibration error: the gap between the agent's stated confidence and its empirical accuracy.
- Revision rate: how often submitted forecasts are revised.
- Materiality flag rate: the fraction of forecasts flagged as material.
- Human override rate: how often a reviewer reverses the agent's forecast.
BLOCK above 15% human override rate or above 40% revision rate. LIMIT above 8% human override rate, above 25% revision rate, below 80% forecast accuracy, or above 20% calibration error.
Attestations and verification
Attestations cover a fixed period of forecast decisions, reviewer assessments, and the latest reliability signals. The report is serialized deterministically and hashed with SHA-256. The verification string format is SIGMODX-FORECAST-[ORG]-[HASH]. Auditors can independently verify the string at /verify.