Model & accuracy

The whole record, not the highlights.

Every signal the model has fired is on this page — won, lost and pushed. The headline numbers are computed from that ledger, so nothing here can be a selection of good weeks.

ROI
+0.00%
0u risked
Units won
+0.0
0 settled
Hit rate
0.0%
0W 0L 0P
Beat the close
0.0%
avg +0.00% CLV
Max drawdown
0.0u
worst run 0L
Open positions
0
avg edge 0.0%

Backtest — not live bets

no history yet

Everything above is the live ledger: signals written when they fired, at the price then available. This panel is something weaker and it is kept apart on purpose — the same model replayed over games that had already been played, priced against each game’s closing number using only the ratings that existed before it. No money was on any of it.

Not enough stored history to replay yet. The model needs a competitor to have played a minimum number of rated games before it will price anything, and that floor is set by the sport rather than by what makes this number look better — forty for baseball, six for football. Until the collector has that much behind it, there is nothing here.

Equity curve

cumulative units

Running ROI

return per unit risked

Early swings are sample size, not skill. The curve only means something once the pick count on the left has run into the hundreds.

Calibration

stated vs actual

Each dot is a probability bucket; the dot area is how many games landed in it. On the dashed line, a stated 70% wins 70% of the time. Above it the model is too cautious, below it too confident.

Return by stated edge

does a bigger edge pay more?
-1%0%+1%2.6–4%0%2.6–4% · 0%0 picks · 0% hit4–6%0%4–6% · 0%0 picks · 0% hit6–9%0%6–9% · 0%0 picks · 0% hit9%+0%9%+ · 0%0 picks · 0% hit

The honest test of a model. Buckets should rise left to right; where they do not, the edge estimate in that band is overstated — and the thin buckets on the right are the ones to trust least.

Closing-line value

distribution of every pick
-101-6-6 · 0-6% to -4%-4-4 · 0-4% to -2%-2-2 · 0-2% to -1%-1-1 · 0-1% to 0%00 · 00% to 1%+1+1 · 01% to 2%+2+2 · 02% to 4%+4+4 · 04% to 6%+6+6 · 06% to 9%

How much better than the closing number each bet was taken at. Mass to the right of zero means the model is consistently early to where the market ends up — the single best early read on whether an edge is real.

Monthly returns

units, by month

Prediction error

how wrong, on average
Brier score
0.0000

0.25 is a coin flip

ATS accuracy
0%

model side vs the spread

Margin error
±0

mean absolute, in points

Total error
±0

mean absolute, in points

Measured across all 0 finished games, not only the ones the model bet. A model that only grades itself on its own picks is grading itself twice.

Units by league

    League breakdown

    0 leagues
    LeaguePicksHit rateROIUnitsAvg CLVShare of P&L

    Open positions

    0 live

    No open positions.

    Ledger

    most recent 50, graded
    FiredLeaguePickGamePriceCloseCLVEdgeStakeResultP&L

    How to read this page

    Why CLV comes first

    Closing-line value measures the price you took against the market’s final number. It stabilises after a few hundred bets, while ROI needs thousands — so a model beating the close is showing an edge long before the profit column proves it.

    Why hit rate is the weakest number

    A 53% hit rate at −110 is profitable; a 58% hit rate on heavy favourites can lose money. Price and stake decide the outcome, which is why the ledger records both and the ROI column is the one that settles arguments.

    What would falsify it

    Flat or negative CLV, a calibration curve bending away from the diagonal, or edge buckets that fail to rise with the stated edge. All three are visible above; none of them are hidden behind a summary.