Win Probability
The number on the bar is a calibrated statistical estimate, published so you can read the match and price it. It is not a guarantee, and it is expected to be wrong a predictable share of the time.
What it reads
The model takes live match state — army value, cash and earnings, unit and structure counts, harvesters, production, power, and kill counts — compares the two sides, and factors in how far into the match you are. An army lead at minute two and the same lead at minute twelve are different claims, and the model treats them differently.
How good is it
Measured out-of-fold, split by match, on a corpus of finished matches:
Uncalibrated
0.9124
0.2148
0.723
0.833
0.1767
Shipped (calibrated)
0.5130
0.1747
0.723
0.831
0.0988
Two things worth understanding in that table.
Accuracy did not change, and that is correct. Calibration is a monotone squash — it cannot reorder anything, so it cannot make the model better at picking winners. It changes how confident the stated number is. Ranking quality (AUC) and calibration quality (ECE, log loss) are separate properties, and only the second one was broken.
The uncalibrated model scored worse than answering 50% every time (log loss 0.69) while being 72% accurate. That is the expensive failure mode: confidently wrong. For a number that prices predictions, overconfidence is not a cosmetic problem — it misprices every stake placed on it.
Why it is not more certain than it is
Published research on RTS win prediction lands around 70–77% accuracy mid-match. A model that looks much more certain than that early is not smarter; it is miscalibrated. If you see 91% four minutes into an even board, that is a bug, not an insight.
High confidence late in a lopsided match is legitimate — by then the state genuinely determines the outcome.
How to read it as a viewer
Near 50% early is honest, not indecisive. Most matches are undetermined at minute three.
Sharp moves follow real state changes: an army trade, a lost expansion, a production swing.
It can be wrong. A 70% reading means the favourite loses roughly three times in ten. Those losses are the model working as specified, not evidence it is broken.
Where it runs
The same model, with the same calibration constant, evaluates in the browser (for the broadcast bar) and in the market service (for pricing). They are kept in step deliberately — a viewer and the book must never see different numbers.
Last updated