Sales forecasting accuracy in 2026 depends on splitting pipeline by source. Warm-sourced and cold-sourced opportunities have materially different win rates, cycle times, and average deal sizes. Forecasting them as a blended pool produces 30-40% inaccuracy. This post is the dual-source forecasting methodology.
The quality differential by source
| Source | Win rate | Cycle length | Avg deal size |
|---|---|---|---|
| Warm-intro (champion job change, board, advisor) | 35-50% | 40% shorter | 20-30% larger |
| Customer referral (passive) | 30-40% | 30% shorter | 15-25% larger |
| Cold cadence reply | 15-25% | baseline | baseline |
| Inbound demo request | 20-30% | 30-40% shorter | 10-20% larger |
The dual-source forecasting model
- Tag every opportunity with source at creation
- Apply source-specific win-rate probabilities to weighted pipeline
- Forecast warm and cold pipeline separately, then combine
- Track forecast accuracy by source quarterly to refine probabilities
What changes operationally
- Pipeline coverage ratio splits by source (warm 2-3x; cold 4-5x)
- Quota allocation considers warm-vs-cold mix per rep
- Deal review focuses on source-appropriate questions (warm: "who's the connector, is the EB engaged?"; cold: "have we identified champion + EB yet?")
- CRO board reporting shows warm-vs-cold contribution explicitly
The forecasting dashboard
Weekly view:
- Weighted pipeline by source (warm vs cold)
- Forecast accuracy trailing 4 quarters by source
- Cycle compression measured per source per segment
- Source mix shift over rolling quarters
Common forecasting mistakes in 2026
- Single blended win rate. Produces 30-40% forecast inaccuracy at Series B+.
- No source tagging on opportunities. Can't apply differential probabilities.
- Ignoring connector quality. Top-quartile connectors produce 60-80% win rates; bottom-quartile produce 20-30%. Connector-credit attribution refines further.
Where to start
For pipeline coverage ratio framework, see pipeline coverage ratio 2026. For the operating model, see warm outreach.
.png)
