◇ Methodology
A hundred on a flat deck against a weak attack and a hundred on a green seamer against the best bowlers in the world go into the record book as the same hundred. Everyone who watches cricket knows they aren't. And neither number tells you whether the runs won a match or padded a total that was already lost. Fine Margins works on the modern Test game, and it's built around those two blind spots: how hard the runs were to score, and how much they mattered. This page explains how it tries to measure both, and how often it gets it wrong.
These two questions don't have the same answer, which is why Fine Margins keeps them apart. Two batters can retire with identical averages having got there in completely different ways. One filled his boots in matches already won or lost. The other made match-defining runs with the conditions and the situation stacked against him. Neither a raw average nor a single line on a scorecard can separate those two careers.
The difficulty-adjusted average works on each innings, reading three things from it: the venue, the strength of the bowling attack faced, and whether the batter was at home or away. From those it estimates the difficulty of that innings, then re-weights the runs accordingly, innings by innings.
The model underneath is more detailed than those three factors suggest. It also learns how each ground tends to play, and how a pitch changes as a match wears on, with every venue allowed to behave differently. The published adjustment distils all of that down to the parts that carry the most weight: venue, attack, and home or away. The rest refines how those are estimated rather than adding a separate lever.
The model is Bayesian. In practice that means two things. It shares information across players and conditions rather than treating each innings in isolation. And it returns a range of plausible values rather than a single figure, which is why every adjusted average on the site carries a credible interval. It also means small samples are handled cautiously: a player with few innings has their adjusted average held close to their raw average, because there isn't yet enough evidence to move it far. The fewer the innings, the less the adjustment. The model only says as much as a player's record allows.
Value is a series of calculations, not a Bayesian model. It starts from the runs a batter actually scored in an innings and scales them by how much those runs mattered. Test cricket has many situations where runs carry weight, and the value model attempts to capture all of them: a first-innings hundred, runs dragged out of a deficit, an occupation of the crease that blocks out a draw, and the many other moments across five days where an innings deserves more credit than its size alone suggests. It leans on three things to judge this: how difficult the match situation was, how much the innings swung the result, and how large a share of the team's runs it accounted for. It also takes some account of conditions, so that being in trouble on a flat pitch is not rewarded as heavily as the same trouble on a difficult one.
What value deliberately doesn't do is punish runs for coming at the wrong time. This is a Test-cricket judgement. A future Fine Margins model for the shorter formats will dock a slow innings, because deliveries are scarce and tempo is everything. Test cricket is different. Time is rarely the binding constraint, and beyond a low score almost any runs help the side. So value lifts the innings that mattered most without penalising the rest. It rewards, it doesn't punish.
Because it is calculated one innings at a time, never pooled or shrunk, value stays steady at any sample size. A career, a single series, or a handful of fourth-innings chases can all be totalled the same way, which is why the splits explorer leans on value when the sample gets too thin for the adjusted average to say much.
The model sees what happened in an innings, not how it happened. It knows the runs, the situation, the conditions and the result, but it has no ball-tracking, no record of plays and misses, no measure of how cleanly a batter struck the ball. A scratchy fifty riding its luck and a flawless one look identical to it. Systems built on ball-by-ball tracking data can see some of this; Fine Margins, for now, cannot.
It also doesn't model form. A career is treated as a body of work rather than a sequence, so hot streaks, slumps and the decline of a player's final years are not captured as such. This is a deliberate choice for the first version rather than an oversight, and one worth a longer explanation another time.
The approach behind both models is to set the structure and let the data fill it in. The choices that are ours are choices about the frame: which factors a model is allowed to see, how the pieces fit together, what counts as a fair reference point. Within that frame, the numbers are not opinions dialled in by hand. They are what the record says once it is read through that structure. A small number of settings, like how strongly the adjusted average rewards difficulty, are tuned by judgement to stay interpretable, and where that is true it is said plainly. But the headline figures are derived from the data.
Every adjusted average comes with a range, not just a single figure. The headline number is the best single estimate; the interval beside it shows where the true value most plausibly sits. A wide range is the model admitting it isn't sure, which happens most with players who haven't batted often. Showing it is the honest thing to do, so the numbers are read with the confidence they deserve and no more.
The numbers are built from ball-by-ball data published by Cricsheet, covering men's Test cricket from the early 2000s to the present. The Bayesian modelling follows the approach set out in Richard McElreath's Statistical Rethinking. Fine Margins is an independent project and is not affiliated with any cricket board or governing body.