EXECUTIVE SUMMARY / THE DECISION IN PLAIN LANGUAGE
A better forecast still needs a warning system.
Evaluated synthetic study · no actual client outcomes
In the stable simulation, the detailed model's 12-week revenue error was 6.5%, versus 19.6% for fixed stage weights. The simpler timing model was slightly more accurate at 6.0%.
What a leader should do next
Pilot the forecast alongside the existing sales process, with clear alerts when conversion rates or deal timing change.
What could change the decision
When the simulated sales process slowed abruptly, all twelve 95% forecast ranges missed. The forecast covers only opportunities already open. It cannot support a firm revenue commitment.
01 / BUSINESS DECISION
Start with the decision.
Create a 4-, 8- and 12-week revenue outlook for opportunities already open at the forecast date. Separate conversion and timing uncertainty from sales commitments. Future opportunities are outside the forecast population.
02 / PRACTICAL IMPLICATION
What the evidence can support.
In the stable synthetic process, 12-week aggregate forecast error is 6.5%, compared with 19.6% for fixed stage weights. A simpler pooled-stage model does slightly better at 6.0%. A sudden conversion slowdown causes severe overprediction and every nominal 95% interval misses. A production pilot would need process-change alerts and a fallback outlook before any reliance on the ranges.
03 / EVIDENCE
Revenue forecasts under changing sales conditions.
Three implemented forecasters are evaluated at twelve held-out forecast dates, over three horizons and three declared scenarios. The study generates 5,760 opportunities per scenario and preserves weekly point-in-time snapshots. No final-period tuning is used.
Interactive results · the executive summary above remains the fixed headline comparison.
SYNTHETIC PIPELINE · 5,760 OPPORTUNITIES PER SCENARIO · 12 FORECAST DATES
The detailed model improves on fixed weights, while the simpler timing model remains a strong comparator. Greater complexity is not a uniform advantage.
| Method | Revenue error | Signed bias | 95% range coverage | Probability error¹ |
|---|---|---|---|---|
| Fixed stage weights | 19.6% | -19.6% | 41.7% | 0.257 |
| Pooled stage model | 6.0% | 0.7% | 100.0% | 0.234 |
| Detailed multistate | 6.5% | -3.0% | 100.0% | 0.228 |
Revenue error is total absolute forecast error divided by total realized revenue (WAPE). Positive bias means overprediction. ¹Probability error is the Brier score; lower is better. Twelve overlapping origins are dependent, so coverage is a descriptive diagnostic with limited precision.
Probability calibration and interval width
| Probability band | Snapshots | Predicted | Observed |
|---|---|---|---|
| 20%–40% | 1464 | 33.7% | 31.8% |
| 40%–60% | 1254 | 48.2% | 47.5% |
| 60%–80% | 40 | 61.3% | 57.5% |
Mean 95% revenue-range width: $355,169. An opportunity may appear at multiple forecast dates; these are forecast snapshots, not independent customers.
04 / DATA
Know where the evidence comes from.
Original synthetic CRM event histories, version synthetic-pipeline-v1. Every opportunity has an entry date, fixed amount and segment, weekly stage/age, and an end-of-week transition. A win books the full amount once; a loss books zero. No real CRM records are included. Open deals are censored at a cutoff and contribute exposure without being labeled lost.
The source includes a data dictionary, fixed configuration and seed, original data generator, observed CSVs, evaluation outputs and source/data hashes. Read the data dictionary ↗
05 / METHODOLOGY
Match the method to the design.
Estimate weekly probabilities of staying in stage, advancing, winning or losing. The primary model conditions on three stages, two segments and three age bands, with Dirichlet smoothing. Propagate these competing transitions to horizon-specific win probabilities. Aggregate 500 posterior predictive draws that share a sampled transition table across opportunities, then simulate individual paths. Compare against the same open pipeline under both baselines.
Evaluation design
- Twelve forecast origins, four weeks apart; 4-, 8- and 12-week outcomes are fully observed.
- Fixed stage weights, pooled-stage competing risks and a Bayesian stage/segment/age model.
- Aggregate revenue error and bias, opportunity probability calibration, and 80%/95% predictive coverage under stable, slowdown and shared-shock conditions.
References motivate the methods. These study-specific implementations do not claim every assumption or theoretical guarantee of the cited work.
06 / LIMITATIONS
Conditions that matter.
- Only transitions strictly before a forecast cutoff enter estimation; the open snapshot contains start-of-week features only.
- Observed past transitions of currently open deals are valid information. Their later outcomes are excluded until their event dates.
- Deal values stay fixed and are recognized fully upon winning; future-created deals, partial recognition and pipeline revisions are excluded.
- The conditional transition process must remain useful. Posterior parameter sharing does not represent an unobserved market-wide shock.
This is a completed synthetic evaluation, not demonstrated CRM accuracy. Twelve overlapping forecast windows provide limited, dependent evidence about interval coverage. Process drift and correlated deals can invalidate the predictive ranges, as the stress tests show. The more detailed model is not uniformly better than the simpler baseline.
07 / CODE & NEXT STEPS
Reproduce the result.
git clone https://github.com/mpgibb/revenue-forecasting.git
cd revenue-forecasting
uv sync --frozen
uv run python -m unittest discover -s tests -v
uv run python study.py
git diff --exit-code -- data resultsAutomated checks and synthetic evaluation are complete; Michael’s independent technical review is pending.