The Daily Lob

Where the tournament stands

The change, and the number behind it

From today this page is about value rather than accuracy. Those are different questions, and the brief has been answering the easier one.

Accuracy asks whose probability sat nearer the outcome. Value asks whether backing the disagreements at the real quoted prices would have made money. A model can be less accurate overall and still find value, because value lives in the few spots where it disagrees most and accuracy is dominated by the many cards where the two lines nearly agree. It can also be more accurate and still lose, because the bookmaker's margin has to be cleared before a correct opinion becomes a profitable one.

So the second question was asked properly, on every scored card this brief has published. Every edition carries the decimal odds actually printed on each card; every following edition carries the result. Joining them gives a backtest on real prices — back the side Elo likes more than the market does, one unit, at the quoted odds. 83 scored cards. Return: -2.39 units. ROI: -2.9%.

The first version of this table was wrong, and how it was wrong is worth more than the table. It joined each result to its price by matching an exact surname string — and the earliest editions do not define the helper that produces those surnames, while several cards print a shorthand (“vdZandschulp”, “M. Perricard”, “Mérida Aguilar”) that the helper never generates. Thirteen of eighty-three scored cards were being dropped, silently, and the number read −18.7%. Joined properly it reads −2.9%. A backtest quietly missing a sixth of its sample is worse than no backtest, so the tool now prints every card that fails to match; today it prints none.

Read the shape, not the best row. A real edge should strengthen as the disagreement grows. This one does not: it is positive at 2 points, negative at 5, negative at 8, then strongly positive at 10 and 15 on twelve bets and five. That non-monotonic pattern is the signature of noise rather than signal, and the two positive tail rows rest on almost nothing — a single long-priced winner moves them by tens of points. The direction and the sample size have to be reported together.

The honest summary is: roughly break-even, slightly behind, across eighty-three bets — which is about what should be expected when a weekly ratings file five days stale argues with a live market, before bookmaker margin of roughly 2–4.5% is cleared. It is not evidence of an edge and it is not evidence there is none. It is the number that decides whether any of this is worth doing, so it goes at the top of the page every day from now, whichever way it points.

Yesterday's brief, scored

Seven cards, seven results, the favourite won four — and the closest price this brief had ever printed came in as flat as it looked.

Three underdogs, and the biggest of them was the biggest price on the page. Emma Navarro beat Karolína Muchová 6-3 7-5 as a 22% underdog against a woman who was 35-9 for the season. Yesterday's card noted that the model liked Navarro materially more than the market did — 28.4% against 22.2%, the only card where it favoured the underdog by a real margin — and that is the one it got right. Anna Kalinskaya beat Elina Svitolina 6-3 2-6 6-3 and the two did not shake hands. Stefanos Tsitsipas beat Jiří Lehečka 2-6 6-1 7-6(3) 6-1 after losing the opener.

Ben Shelton won in four, in 3h22, finishing after two in the morning — despite a head-to-head that read 4-0 to Denis Shapovalov, which yesterday's card called the sharpest contradiction it had carried. It was right that the contradiction was sharp and wrong about which way it would fall. One correction: sources differ on the third set, given as 6-2 by one and 6-3 by another; the brief could not settle it and prints neither as certain.

The closest price in twenty editions did nothing interesting. Etcheverry v Navone was priced 50.6/49.4 and Elo agreed to within half a point; Etcheverry won 6-4 6-4 6-3 in straight sets. Under yesterday's ordering that was the lead card. Under today's it would not have been carded at all — which is the change, stated as plainly as it can be.

Fifteen days. Selected matches played to a finish have gone to the favourite 69 times in 118.

Elo v Vegas

Elo was closer on 5 of seven. Across fourteen days it stands at 39 of 84.

This is the accuracy line, and it is kept because dropping a measure the day it stops flattering you is bad practice. But it is no longer the headline. Five of seven is a good day and the running figure is still under half — and the backtest above is the reason that matters less than it looks: the cards where the two lines nearly agree dominate this count and are worth nothing either way.

The ratings are five days old and have still seen no match of this tournament. Both files remain stamped 31 August. Everything on this page is computed from a world in which the US Open has not started — which is most visible on card 01, where the model has not seen van de Zandschulp beat Medvedev, Hurkacz and de Minaur, and on card 06, where it has not seen Rybakina's ankle.

The deep-run flag

Added on 30 August at a reader's suggestion: be wary of a player coming off a semifinal or final at their previous tournament, within about ten days. It is run as a tracked experiment, not an opinion. Five settled cases so far; they have won three, and their own prices implied 2.6. That is running against the thesis — flagged players have so far done slightly better than the market expected, not worse. Five is far too few to conclude anything, and the count is reported instead of a verdict.

The test is not whether flagged players lose — a favourite coming off a final should still win most of the time. It is whether they win less often than their own market price implied. If they do not, the flag is decoration and it comes off the page.

Today adds Elise Mertens — Monterrey runner-up, seven days ago, five matches in six days. She is also the side the value board says to back, at +240. The two signals point in opposite directions at the same player, and the page is publishing both rather than quietly dropping the inconvenient one.