National League week one results 2026/27: every result, and the model scores we never published

What we are publishing, and why it is a weak week

The National League's 2026/27 season opened on Saturday 8 August 2026 with 12 fixtures at 3.00pm. Our model scored all 12. It showed none of them to anybody.

This piece publishes those 12 scores for the first time, alongside the results. Of the 12, the result the model leaned towards came in 5 times.

That is a weak week and we are not going to present it as anything else. It is also not our record, for a reason that matters more than the number. Not one of those 12 fixtures was ever released. Every one of them carried an insufficient-data flag, none appeared on a board, and none was shown to a single user at any point. What follows is a look inside the engine, not a reveal of something that was live.

We are publishing it because the season is a day old and there is no record here to protect. Showing a weak week costs nothing in August and a great deal in April, so it goes out in August.

Every National League week one result, with the score we never published

Scores are out of 10. "Factors held" counts how many of the model's 8 data factors had any 2026/27 evidence behind them on the morning of the match.

The day itself finished 5 home wins, 3 draws and 4 away wins.

Five from twelve, and the arithmetic underneath it

The model leaned home in all 12 fixtures. So its five correct calls are the day's five home wins, and they could not have been anything else. When a model says the same thing twelve times, it does not beat the base rate or lose to it. It becomes the base rate.

This is not a quirk of opening weekend. Across our settled record of 192 fixtures from last spring, the model leaned home in 175 of them. The direction of the lean is close to a constant. It is not where the information sits.

The information is supposed to sit in the conviction score, and last Saturday the score did nothing at all:

• The five fixtures the model got right carried an average score of 3.4.
• The seven it got wrong carried an average score of 3.3.
• The two highest scores on the card, 7.3 and 5.8, were both wrong.
• The lowest scores on the card, three fixtures at 0.0, went 1 from 3, which is roughly what the whole card did.

A conviction score exists to separate the fixtures worth reading from the fixtures that are noise. On 8 August it separated nothing. A gap of 0.1 between the right calls and the wrong calls is not a signal, it is a rounding artefact.

What the model actually held on those 12 fixtures

Pulled straight from the scoring engine: those 12 fixtures held an average of 0.8 of the 8 factors. Our settled spring record averaged 6.8.

Ten of the 12 rested on head-to-head record alone. Two rested on nothing whatsoever, because the clubs have no recent meetings to read: AFC Hornchurch v Kidderminster Harriers, and Carlisle United v Worthing, who came up this summer.

And one fixture shows the flaw more clearly than any argument could. Boreham Wood v Tamworth scored 0.0 with a head-to-head factor present. It read "won 0 of the last 2 meetings", which scores zero. Carlisle v Worthing also scored 0.0, with no factor at all. On the page they print the same number. One of them means "the evidence points away from the home side" and the other means "there is no evidence". A score cannot tell you which, and neither can a reader.

Why a head-to-head-only score is not a conviction score

Here is the mechanism, and it is the single most important thing on this page.

When the model has a full factor set, the eight factors carry these weights:

Head-to-head is the joint-smallest input in the model. It is designed to be 8% of a conviction score.

When only one factor is available, the model renormalises, and that factor's weight becomes 1. So on opening weekend, head-to-head carried 100% of a number built to let it carry 8%. The score still prints to one decimal place and still looks like every other score we produce. It is a different quantity wearing the same badge.

Scunthorpe United v Yeovil Town is the clean example. It scored 7.3, which in a normal week is a high number. The entire 7.3 came from "Scunthorpe won 3 of the last 5 meetings". Nothing about either squad this summer, nothing about form, nothing about shots, nothing about absences. It finished 1-1.

Why a fixture stays unrated until both clubs have played four league matches

Two separate gates govern what leaves the building, and last weekend a fixture had to fail only one of them.

Gate one: the four-match rule. A fixture is flagged insufficient data until both clubs have four league matches of the current season behind them. Six of the eight factors are read from this season and only this season, so before four matches they are either empty or built on a sample small enough that one deflected goal moves them. Four is where recent form stops being a coin, where goals and clean sheets per game stop swinging by half a goal a week, and where a league table starts to mean something rather than reflecting who has played whom.

Gate two: the release threshold. We only put a fixture on a board at 7.0 and above. Everything below that is filtered out before anyone sees it.

Scunthorpe v Yeovil cleared gate two at 7.3 and failed gate one. It stayed in, and that is the gate doing exactly what it exists to do. A 7.3 built on one input would have gone out looking identical to a 7.3 built on eight, and the number would have been worth less from that day onwards. The full reasoning is in why we don't score fixtures until four games into the season.

The model scores everything, always. The gates decide what is published. Those are different jobs and it is worth keeping them apart in your head.

What our settled record says, and what it does not

For contrast, this is what the conviction score does when it has its evidence. Across 192 settled fixtures between 14 April and 25 May 2026:

Update, 10 September 2026. That table is the spring window only. Across 17 settled released picks between 31 August and 5 September 2026 the picks at 8.0 and above landed 2 times in 6 and the picks below 8.0 landed 6 times in 11, so the staircase has run backwards in the new season. Our released record across both windows is 27 from 39 released picks, 14 April to 5 September 2026.

That is a score sorting fixtures, which is the whole point of having one. The band we actually release from, 7.0 and above, landed 19 times in 22 across the released window of 14 April to 2 May 2026, which is 86% with a 95% confidence interval of roughly 65% to 97%. Twenty-two fixtures is a small sample and that interval is wide, so treat it as the best estimate we hold rather than a settled figure. It will move as 2026/27 settles, and we will move it in public when it does.

Two honest cautions on reading last weekend against that table.

First, 12 fixtures would not settle this question in either direction. Had the model gone 9 from 12 on the same 0.8 factors, it would have proved nothing either, and we would have said so in the same words.

Second, the staircase above is built on the full factor set. Last weekend's scores were not comparable numbers, so "the 7.3 drew" is not evidence against the 7.0-and-above band. It is evidence for the four-match gate, which is the argument this piece is making.

When the first rated National League fixtures arrive

The National League started on 8 August, so both clubs in a fixture reach four league matches around the end of August. The first National League fixtures to lose the insufficient-data flag should appear in late August, and only those where both clubs qualify. The board may still be empty after that, because a rated fixture still has to reach 7.0.

The Championship, League One and League Two start on 14 and 15 August, which puts their first rated fixtures in early September.

Until then the honest answer to "what does the model say about this weekend?" is that it says very little, and we would rather write that sentence than manufacture a number.

Lower League Analyst is a data research tool, not a betting product.