Two weeks into every quarter, Tallowtree's reps commit to a number. Leadership asked for a second read from the CRM's own history to sit next to it.
For every open deal on day 14 of the quarter, the model reads two things from the CRM: the furthest stage the deal has reached, and how many of three buyer-evidence fields are filled (budget confirmed, economic buyer met, mutual close plan). It looks up how often deals in that same position won within the quarter, in earlier quarters only, and multiplies by the deal's annual value. It then adds the average value of deals that were created after day 14 and still won in the same quarter, over the last four quarters.
Furthest stage, because deals get moved backward when a rep tidies up, and the furthest point tells you more than where a deal sits today. Buyer evidence, because a deal at Proposal with a confirmed budget and a met economic buyer is a different deal from one with neither.
The rates it uses for Q4 2026, learned from every quarter before it:
| Furthest stage | No evidence | 1 field | 2 fields | All 3 |
|---|---|---|---|---|
| Solution | 27% 59 deals | 16% 91 deals | 36% 33 deals | 23% 0 deals |
| Proposal | 33% 18 deals | 46% 37 deals | 71% 41 deals | 67% 15 deals |
| Negotiation | 59% 13 deals | 53% 40 deals | 65% 34 deals | 59% 7 deals |
Gray cells have fewer than 15 deals behind them, so the model uses that stage's overall rate instead. Some neighboring cells don't rise in order. That's real noise from small groups. I leave it in, because smoothing it would hide how thin the data is.
Each quarter's model only saw what was in the CRM on its own day 14, with rates learned from earlier quarters. Nothing from the quarter being called leaks in.
Won Model, day 14 Reps' commit, day 14 Model range, 80%
| Quarter | Won | Model | Miss | Commit | Miss |
|---|---|---|---|---|---|
| Q1 2025 | $606K | $614K | 1% | $425K | -30% |
| Q2 2025 | $860K | $700K | -19% | $541K | -37% |
| Q3 2025 | $945K | $1.01M | 7% | $621K | -34% |
| Q4 2025 | $960K | $789K | -18% | $688K | -28% |
| Q1 2026 | $768K | $776K | 1% | $766K | 0% |
| Q2 2026 | $896K | $666K | -26% | $589K | -34% |
| Q3 2026 | $752K | $787K | 5% | $727K | -3% |
A forecast can be right in total and wrong deal by deal. This check groups every back-tested deal by the chance the model gave it and counts how many actually won.
| Model's chance | Deals | Average chance given | Share that won |
|---|---|---|---|
| 0% to 5% | 315 | 4% | 3% |
| 5% to 20% | 248 | 11% | 13% |
| 20% to 40% | 55 | 30% | 35% |
| 40% to 60% | 56 | 51% | 45% |
| 60% to 100% | 78 | 70% | 56% |
The low bands are close. The top band is too generous: deals the model gave 60% or more won less often than that. On a client's data I'd find out why before trusting the top of the range. Deal size is the first thing I'd check.
| Furthest stage | Open deals | Annual value | Model expects |
|---|---|---|---|
| Solution | 20 | $564K | $134K |
| Proposal | 7 | $235K | $145K |
| Negotiation | 8 | $212K | $116K |
| Discovery and Qualified | 68 | $184K | |
| Not created yet | $62K | ||
| Total | 103 | $642K |
The range comes from playing the quarter 4,000 times: each open deal wins or loses at its own rate, plus one of the last four quarters' late wins. In 80% of those runs the quarter landed between $454K and $850K.
The model is a second opinion beside the reps' call, and the useful part is where the two disagree. Those are the deals to talk about in the forecast meeting.
I build and back-test this on your CRM data. Fixed price $2,500, scoped in writing before I start. Hire me on Upwork.