Alpha you can extract

The Traditional Quant Story

Every fundamental research process begins with a problem of scale. The investable universe runs to tens of thousands of companies; and for decades quantitative filters have narrowed it down.

The logic is honest and appealing. The raw data is there — EPS estimates, analyst revisions, dividend histories, CFROI. Somewhere in that data sits alpha: a faint signal specifying which companies are worth a human’s attention. So quants engineer factors — a consensus revision measure here, an analyst-direction ratio there, a CFROI trend — each one a crystallised piece of human intuition about where value hides. Each factor is ranked across the universe, the ranks blended together with a set of weights. The top five hundred fall out of the bottom of the funnel, and the research begins.

It is a clean pipeline, and it rests on three quiet assumptions: that the alpha is in the base data, that the hand-built factors preserve it, and that the final process exposes it. The first may well be true. But the factors and the formula are human guesses about the shape of the signal — and nothing guarantees that a difference of EPS revisions or a weighted rank-sum is the right lens through which to see it. The alpha can be present and the method still blind to it.

The AlphaFold Story

For twenty-five years, the structure of a folded protein was one of biology’s great unsolved problems. The relationship was known to exist: a protein’s three-dimensional shape is determined by its amino acid sequence. The answer was in the sequence. CASP, the field’s biennial blind test, set the world’s best minds — mathematicians, physicists, biologists, programmers — against that sequence-to-structure problem, and for a quarter of a century they engineered features, energy functions, and heuristics that captured their best understanding of how folding worked. They were brilliant, and they fell short. The information was there; the human-designed method of extracting it was not equal to the task.

Then, in 2020, AlphaFold2 — at heart a deep neural network — predicted structures to roughly experimental accuracy, and the organisers declared protein folding largely solved. It won its creators the 2024 Nobel Prize in Chemistry and is widely regarded as the applied-mathematics breakthrough of the decade. What changed was not the input. The sequence had always held the answer. What changed was that a learning system was allowed to discover the mapping for itself, rather than have humans hand-build it.