A bad call is only wasted if the process that produced it stays hidden. The word itself borrows from medicine: as the Cambridge Dictionary notes, the prefix "post-" simply marks what happens after an event, and a post-mortem is the examination that follows it.https://dictionary.cambridge.org/dictionary/english/post In investing, that examination has one job — to find out whether the decision was bad, or whether a good decision simply met an unfavourable outcome.
That distinction is the whole discipline. A sound thesis can lose money. A reckless thesis can double. If a team grades itself on results alone, it trains itself to repeat lucky recklessness and abandon sound judgment. A structured investment decision review process grades the reasoning, the information available at the time, and the execution — then lets the outcome inform, but not dominate, the verdict.
What follows is a working method: what to capture before the review, how to run it, and where these reviews routinely fail.
What is a post-mortem, and why does outcome bias break the obvious version?
A post-mortem is a written, scheduled review of a completed decision that reconstructs what the decision-maker knew at the time and judges the process against that record. It is not a performance chat and not a search for someone to blame.
The obvious version — "it lost money, so the call was wrong" — fails because of outcome bias. Outcomes in markets carry a large random component. A concentrated position can be wiped out by one headline no analyst could have priced. A diversified, well-hedged book can still lose in a drawdown. Judging the decision by the result conflates skill with variance, and the lesson drawn is noise. We covered a connected angle in What does out-of-sample mean in a forecasting paper?.
The fix is procedural, not psychological. Write down, at the time of the decision, the thesis, the expected holding period, the base case and the downside case, and what evidence would falsify the thesis. Then the review compares reality to that record — not to memory, which conveniently rewrites itself after the fact.
What should be captured at decision time?
The review is only as good as the contemporaneous record. A decision journal needs a small, fixed set of fields, filled in before the position is taken:
- The thesis. Why this asset, why now, in one or two sentences a colleague could repeat back.
- The base case and the bear case. Not a point forecast — a range, with the reasoning for each.
- The falsifier. What specific development would prove the thesis wrong, and by when.
- Sizing rationale. Why this position size relative to the portfolio and the conviction level.
- What was deliberately not known. Named gaps — the data not pulled, the management question not asked.
The last field matters more than it looks. At review time, the gap that hurt is obvious; the discipline is recording which gaps were accepted in advance. That is the difference between a known limitation and an oversight.
How should the review meeting actually run?
A useful review has four stages, in order, and the order is the point.
- Reconstruct. Read the original decision record aloud. No updating it to match hindsight.
- Separate process from outcome. Score the decision on the information available then: Was the thesis falsifiable? Was the sizing consistent with conviction? Were costs and liquidity considered? Readers who work on model validation will recognise the pattern — the same separation of in-sample reasoning from out-of-sample result that out-of-sample testing demands of a forecasting paper.
- Classify. Every decision lands in one of four cells: good process with a good outcome, good process with a bad outcome, bad process with a good outcome, bad process with a bad outcome. The cells to worry about are the off-diagonal ones. A bad process that paid is the most dangerous entry in the ledger.
- Extract one transferable rule. Not five. One concrete, checkable change — "size by conviction tier, documented before entry" — beats a page of sentiment.
Keep the meeting short and factual. The chair's only real job is to stop the conversation sliding from "what did we know" to "who is at fault". Blame ends information flow; the whole point of the exercise is to make it safe to surface what went wrong.
What does a good post-mortem culture look like across a team?
One review changes little. Culture is the pattern over many reviews, and three features mark it.
Reviews happen after wins too. A post-mortem after a large gain asks whether the process deserves the credit, or whether variance does. Teams that only review losses systematically over-weight the lessons of bad luck and under-weight the lessons of good luck.
The ledger is visible. Classifications, extracted rules, and whether past rules were followed should be tracked where the whole team can see them. A rule nobody checks is a rule nobody follows.
Rules get retired. A process rule extracted from one bad call is a hypothesis, not a law. When a rule stops earning its keep — when it blocks reasonable decisions or contradicts newer evidence — retire it with the same documentation it arrived with.
There is also a modelling parallel worth naming. The same discipline that governs a trading review governs a model evaluation: fix the procedure before you see the result, then judge against it. Work on walk-forward validation makes the same argument for backtests — what the method fixes and what it leaves open must be stated in advance, not negotiated after the numbers arrive. Readers following this should also see Walk-forward validation: what it fixes, what it doesn't.
Where do post-mortems fail in practice?
Three failure modes account for most of the waste.
hindsight dressed as review. If the meeting starts from the outcome and works backwards, it will find a tidy story. The contemporaneous record is the antidote; without one, the review is fiction with a timestamp.
Too many lessons. A review that ends with a list of nine action items ends with zero. Attention is the scarce resource; one enforced rule per review compounds, nine documented intentions do not.
Silent survivorship. Teams review the positions they still hold and the calls they still remember. The quiet exits — positions closed early, ideas never funded — carry information too, and they are systematically missing from the record. This is the same selection problem that distorts fund databases when survivorship bias inflates the visible track record; the sample you see is not the sample that existed.
Limitations: what a post-mortem cannot establish
Be honest about the ceiling of the method.
A post-mortem cannot convert a losing decision into a winning process. Some losses are the cost of a sound approach; the review's job is to confirm that, not to repair it. It also cannot settle counterfactuals with certainty — "what would have happened if" has no data. And it cannot audit judgment it cannot see: a review of the written record evaluates the reasoning that was documented, which is one reason the decision-time capture matters so much.
Finally, classification itself is a judgment call, and reviewers bring their own biases to it. Two people can score the same decision differently. The mitigation is not a cleverer rubric but a stable one — fixed criteria, applied the same way each time, with the disagreements recorded rather than smoothed over.
Practical steps: a minimal starting kit
For a reader starting from nothing, the smallest workable version fits on one page.
- Open a decision record for every position above a chosen size, with thesis, cases, falsifier, and sizing rationale.
- Schedule the review at exit, not "eventually". A date on the calendar is the difference between a review and an intention.
- Score process, classify into the four cells, extract one rule.
- Keep the ledger visible and revisit the rules twice a year.
None of this predicts markets, and that is the point. The exercise treats forecasting as what it is — a distribution of possible outcomes, not a destiny — and invests instead in the one thing a decision-maker controls: the quality of the process, measured honestly, over time.
