The same works by decade. Almost everything on the method side predates the computing power that made it urgent, and the empty 1980s column is not an oversight.
Brier, 1950
Four pages that founded the whole field. Still the clearest statement of why a probability forecast needs a proper score.
Murphy, 1973
The decomposition into reliability, resolution and uncertainty. Explains why a forecaster can be perfectly calibrated and useless.
Gneiting and Raftery, 2007
The definitive treatment of proper scoring rules. Read this before designing any scoring scheme.
Tetlock and Gardner, Superforecasting
The empirical case that forecasting skill exists, is measurable and is trainable. Popular, and the underlying tournament data is serious.
López de Prado, Advances in Financial Machine Learning
Purging, embargoing, combinatorial cross-validation and the false strategy theorem. The most useful single volume on this list.
Bailey and López de Prado, 2014
Deflating a performance statistic for the number of trials. Short and unanswerable.
Harvey, Liu and Zhu, 2016
Why the conventional significance threshold is far too generous once you count how many things have been tried.
Ioannidis, 2005
Why most published research findings are false. Written about medicine, applies everywhere.
Jaynes, Probability Theory: The Logic of Science
Probability as extended logic. Opinionated, occasionally combative, and the best book on the shelf.
Jaynes, 1957
The paper that connects information theory to statistical mechanics. Where maximum entropy comes from.
Gelman et al., Bayesian Data Analysis
The working reference. Less philosophy, more of what to actually do.
Cover and Thomas, Elements of Information Theory
Entropy, mutual information and the coding view of prediction.
Marchenko and Pastur, 1967
The eigenvalue distribution of a random covariance matrix. The result that tells you how much of your correlation matrix is noise.
Laloux, Cizeau, Bouchaud and Potters, 1999
Applies it to real data and finds that most of the structure is not there. Four pages, enormous consequences.
Bouchaud and Potters, Theory of Financial Risk
The econophysics programme set out properly, including where it does not work.
Amari, Information Geometry and Its Applications
Distributions as points on a manifold. Hard going and worth it.
Coles, An Introduction to Statistical Modeling of Extreme Values
Block maxima, threshold exceedances, and why the estimate that matters rests on the fewest points.
Hamilton, Time Series Analysis
The standard reference. Regime switching, unit roots, structural breaks.
Taleb, Statistical Consequences of Fat Tails
Where standard statistical intuition fails outside thin tails. Abrasive, technically substantial.
Kahneman, Sibony and Sunstein, Noise
Variability in human judgement as distinct from bias. Directly relevant to any forecast with a person in it.
Schelling, The Strategy of Conflict
Focal points, credible commitment and coordination without communication. Still unmatched.
Fudenberg and Tirole, Game Theory
The technical reference for signalling and repeated games.
Surowiecki, The Wisdom of Crowds
Popular, and correct about the independence condition that almost everyone quoting it ignores.
Meehl, 1954
Clinical against statistical prediction. Seventy years old and the finding has never been reversed.