I Wrote Down My Predictions for a Year. I Had Remembered Being Right.

For a year I wrote down predictions — work outcomes, small personal ones, a few public — with a date and a confidence. Forty-one of them. At the end I did two things: scored them, and separately wrote down what I remembered having predicted, before looking.

The scores were unremarkable and roughly calibrated. Of the things I called seventy percent likely, about two-thirds happened, which is fine.

The memory exercise was not fine. Of the eleven predictions that turned out wrong, I remembered eight as either right or as things I had been unsure about. In four cases I remembered having explicitly noted the risk that in fact materialised — and the written prediction contained no such note. I had not just forgotten being wrong; I had constructed a memory of having partially anticipated the outcome.

This is hindsight bias and I knew the term before I ran the exercise, and knowing it provided precisely no protection, which is the subject of a different piece. What I want to describe here is how convincing the false memories were. They did not feel like guesses. They felt like recall, with detail — I could picture the conversation in which I had raised the concern. The written record says otherwise and I trust the record, and I still cannot make the memory feel wrong.

The practical consequence is that I have stopped trusting my sense of my own judgement entirely, in the specific sense of “I usually see these things coming.” I have no evidence for that claim about myself that is not generated by the same process that produced eight false memories.

What I do now is narrow and cheap. Any decision that matters gets three lines written at the time: what I expect to happen, what would surprise me, and how confident I am. It takes ninety seconds. The point is not to be better calibrated, though that may happen. The point is to have a record that is not subject to revision, so that when I later think “I always thought that was risky,” I can check.

Two smaller observations from the year. My confidence was better calibrated on things I could not influence than on things I could — for work outcomes I was systematically overconfident, and for external events roughly accurate. And the predictions I was most certain about were disproportionately the wrong ones, which is a very common finding and which I had assumed did not apply to me.

The uncomfortable implication, which I have not resolved: if my memory of my own predictions is unreliable at this scale over one year, then my sense of my own competence — which is built from exactly these memories — is built on a corrupted record. The only fix I know of is to write things down, and writing everything down is not possible, so the honest position is that I know a little about my judgement in the areas I have measured and essentially nothing about the rest.