The Graveyard.

Everyone publishes their wins. Nobody publishes their denominator.

That asymmetry is survivorship bias applied to research itself, and it is the reason most published backtests cannot be trusted: you are looking at the best row of a table you were never shown. This page is the table.

I run a systematic book on Indian markets and a research process that kills most of what it touches. The kills are not the waste product of that process; they are most of what the process knows. A strategy that survived 143 funerals means something different from a strategy that survived none.

Six graves to start with.

Each card links to its row. The other 79 are below, and the ledger search finds anything.

Five times the test itself was wrong, and I said so in public.

Kills prove discipline. Retractions prove the discipline points at me too.

Read as a distribution, the causes of death are the finding.

60% died as no-signal. 51 of 85 published hypotheses produced nothing once artifacts and controls were applied. That is the base rate this work actually runs at, worth stating plainly because most published research implies the opposite ratio.

13% died because my own test was broken. 11 rows are infra-invalid: look-ahead in the plumbing, stale marks, a statistic computed on a cost-shifted series, a calendar artifact. That is the number I did not expect when I started counting, and it is the strongest argument for keeping the ledger at all, because an error rate you cannot see is an error rate you cannot reduce.

13% were real and unharvestable. 11 rows are cost-wall: genuine gross edge, dead net of honest costs. The most expensive kind of dead end, because they survive every statistical check and fail only where money changes hands.

The rest divide into premium the mechanism hands back (6), era-dependence (5), and one structural wall where the market simply says no.

Recurring failure modes, in the order they have cost me most: a field only knowable after the decision timestamp; a universe screened today and run backwards; costs assumed rather than measured; a statistic computed on a series that was already smoothed; and a result found by search but reported as though it were found by hypothesis.

Not the same kind of evidence.

Of the 85 published rows, 29 carry a stated numeric bar fixed before the test ran. Those are confirmatory: the result either cleared a threshold set in advance or it did not. The remaining 56 were exploratory, run to see what was there, with the verdict recorded afterwards.

Exploratory work is legitimate and necessary, and it is how you find out which questions are worth asking. It is not evidence at the same strength, and treating it as though it were is precisely the error this ledger exists to document. The preregistered bar column is the marker: where it holds a number, the test was confirmatory. Everything appended to the forward file from launch onward is confirmatory by construction, because the bar is committed before the result exists.

Three, stated because a reader would find them anyway.

Rows record what was tested and how it died, not why I believed it would work. The rationale is the most useful field for someone learning to do this, and it is the one the backlog lacks. Autopsies carry it and are being added over time.

The published set is not the complete set. Rows sitting too close to live book construction are withheld, counted and stated. More importantly, configurations discarded before this ledger existed were never recorded at all, so the true denominator is higher than the number shown and is not recoverable.

Configuration counts are deliberately conservative. The census takes the first number in each row's configs field, which undercounts compound entries. The recorded total is a floor rather than an estimate.

Timestamps, stated precisely.

"Preregistered before results" is only as good as its clock, so here is exactly what the clock shows and exactly where it stops short.

Every preregistration file in the research repo was committed as its own commit, before any result was appended to it. There is no case where a bar and its verdict landed together. The gaps, though, are minutes to hours rather than days, for the plain reason that most of these tests take minutes to run: median twelve minutes between the bar going in and the file next being touched. Anyone treating a twelve-minute gap as strong evidence of good faith should not, and I would rather say so than let it be discovered.

What those short follow-up commits actually contain is the part that carries more weight than the gap. They are protocol amendments that declare their own position relative to the output, in the file, at the time. One reads "declared after Phase-1 output, before any rescoring." Another opens with "epistemic status declared before results: this is a same-data follow-up on a result already seen, so directional confirmation here is not independent evidence." Writing down that your own next test will not count as independent evidence, before running it, is a harder discipline than preregistration and it is the one I would point at.

The limitation is real: that repository is private, because the same files contain live book construction, so you cannot check any of this directly. What exists externally are anchors for the flagship rows: an SSRN working paper, dated public posts, and a retraction trail published the same day the errors were found. From launch onward the situation inverts: new preregistrations append to the public forward file before their results exist, and that git history is a clock anyone can audit without taking my word for anything.

Every major kill-family was also cross-checked against the academic literature: 16 families, 7 confirmed by independent published work, 5 scoped as dead-at-my-venue but alive elsewhere, 4 flagged for bounded re-trials, and zero cases of the literature contradicting a component that survived.

12 hypotheses are alive in some form: deployed, paper-gated, or parked as candidates. Their construction does not publish. The ledger is about the dead; the survivors are the book.

The method is free.

Write the bar before the test. Count everything you run. Classify the deaths. Park with named re-open conditions instead of deleting. Retract in public when the test itself was wrong. The full method, with the table format to steal: github.com/Finance-broski/graveyard.

Corrections welcome: measured disagreements are the good kind. Open an issue with your numbers and method; rows change when the evidence does, and the git history keeps every prior version. If you want this class of scrutiny on your own backtest: Bias Check, fixed price, written verdict.