We take real projects that are partway done and try to predict two things: the total cost when the project finishes (the jargon for this is "EAC"), and which time-period it will finish in. We run many forecasting methods on 107 real projects and measure how close each one gets. Hover over any bar, dot, label or grid cell for a plain-English explanation.
Cost error %
How far the predicted final cost was from the real final cost, as a percentage. 3% means the forecast was within 3% of the actual bill. Lower is better.
Finish-date error
How many time-periods the predicted finish was off by. 0.7 means it was less than one period out. Lower is better.
"Typical" (median)
We report the middle project (half do better, half do worse), so one unusual project does not skew the result.
Completion level
How far into the project we were when we made the forecast: 25% is very early (hard, little information), 75% is nearly done (easier). "Average" blends all four.
Project length
Short to very long. Longer projects are harder to forecast, so we also break the results down by length.
"base" vs "sft", and the "teacher"
A base model is a small AI straight out of the box. An sft model is that same small AI after we taught it, using worked examples from a big, expensive teacher AI (shown here as agent_deepseek...). The whole goal is to make a small, cheap model forecast almost as well as the big one.
Headline: accuracy by method (ordered best-overall: cost-rank + schedule-rank)
Each row is one forecasting method. The left bars show how accurate its cost forecast is; the right bars show how accurate its finish-date forecast is. Shorter bars are better. Methods are sorted with the best all-rounder (good at both) at the top. Pick a point in the project's life below, or choose "Average" to combine them.
Every dot is one method. Further left = better cost forecasts; further down = better finish-date forecasts, so the bottom-left corner is best. The dashed line links the "best trade-offs" - methods where you cannot get better at one thing without getting worse at the other. The big diamond is the best all-rounder. Hover any dot for details.
completion level:
Accuracy by project length (darker = worse; rows best-overall first)
The same methods (rows, best at the top) split by how long the project runs (columns). Colour shows the error: pale = accurate, dark red = inaccurate. This reveals which methods stay reliable on long projects (usually the hardest) and which fall apart. Left grid = cost error, right grid = finish-date error.