Verified bibliography
This is not a list of interesting reading: that already exists, and better done. It is the literature behind each concrete decision in the engine, with what was adopted, what was evaluated and rejected, and why.
Entry rule: no reference is admitted without a DOI or a stable link that has been checked, and without saying which concrete decision it supports. No third-hand summaries.
And it is not a promise: our tests resolve every link on every run. If one returns a 404, the whole project fails. “Verified on such-and-such a date” is a claim; checking it every time is a verification.
Bailey, D. H. & López de Prado, M. (2014) · The Journal of Portfolio Management, 40(5), 94–107
What for: Correcting the Sharpe ratio for the number of times you searched. It is the backbone of the overfitting analysis.
Bailey, D. H. & López de Prado, M. (2012) · Journal of Risk, 15(2), 3–44
What for: Probabilistic Sharpe ratio, and the minimum length a backtest needs before it can claim anything.
Bailey, D. H., Borwein, J. M., López de Prado, M. & Zhu, Q. J. (2017) · Journal of Computational Finance, 20(4), 39–69
What for: The central overfitting test. It led us to the engine’s most important correction: the original threshold would have disqualified half of all clean bots by pure chance.
Bailey, Borwein, López de Prado & Zhu (2014) · Notices of the American Mathematical Society, 61(5)
What for: Formalises how many trades are needed before a good Sharpe ratio means anything.
Brown, R. L., Durbin, J. & Evans, J. M. (1975) · Journal of the Royal Statistical Society, Series B, 37(2), 149–192
10.1111/j.2517-6161.1975.tb01532.x
What for: Detecting whether a bot has stopped behaving the way it used to. Used as an audit question, never as an entry filter.
Bell, R. J. & Dean, P. (1970) · Discussions of the Faraday Society, 50, 55–61
What for: From the physics of glass: how many configurations are genuinely distinct when many of them are correlated with one another.
Twice, while this document was being built, a list of supposed co-authors arrived with attributions that did not survive checking: two authors with no traceable co-authorship, and a third who was indeed a co-author, but of a different paper than the one attributed to them.
A document that exists to provide backing cannot afford a single unverified entry. It would be worse than not having the document at all.