Do decision journals actually work?
If you go looking for evidence that decision journals work, you'll quickly meet a statistic: journaling your decisions improves forecasting accuracy by around 19%, per a study in Behavioral Science & Policy. I went looking for that study, because I wanted to cite it. I couldn't find it. Not in the journal, not in the literature, not anywhere except the blog posts citing each other. Maybe it exists somewhere I couldn't reach, but a number you can't trace to a primary source is a number you shouldn't repeat, so let's set it aside.
What I can offer instead is the honest version: to my knowledge there is no large randomized trial of decision journaling as a complete practice. What exists is evidence for the pieces: strong for the failure it fixes and for the feedback loop it copies, thinner for the forecasting-training results, which a recent reanalysis has put back in question. I find the chain persuasive; you can judge for yourself.
You can't trust your memory of your own judgment
In 1972, before Nixon's trips to China and the USSR, Fischhoff and Beyth asked people to predict specific outcomes: would Nixon meet Mao, would the US recognize China. Months later, they asked the same people to recall their predictions. The recalled predictions had moved toward what actually happened, and subjects were sure those were what they'd said. Hindsight bias is one of the most replicated findings in judgment research.
That matters for learning, because any attempt to improve your judgment by remembering how past calls went is corrupted at the source. The feedback loop runs through an archive that rewrites itself. A written prediction, timestamped before the outcome, is a cheap fix, and it's the one version of your judgment that can't be rewritten later.
Your confidence is miscalibrated
Ask people for ranges they're 90% sure contain the true answer — the length of the Nile, a company's revenue — and the true value should fall outside the range one time in ten. Across decades of replications with students, managers, and physicians, it falls outside four to six times in ten. Russo and Schoemaker, writing in Sloan Management Review, ran versions of this with over two thousand professionals and got surprise rates in that range; they were badly overconfident on questions from their own industry too. Alpert and Raiffa had asked for 98% ranges a decade earlier, where the truth should fall outside twice in a hundred; it fell outside about four times in ten.
So the raw material a decision journal collects — "how sure was I, versus how did it go" — is measuring something that is, for most of us, broken and invisible without measurement.
Calibration improves with scored feedback
Lichtenstein and Fischhoff showed in 1980 that calibration training works: give people rounds of confidence judgments followed by feedback on their actual hit rates, and their stated confidence starts tracking reality, with almost all of the improvement arriving after the first feedback session. It holds outside the lab too. US forecasters' probability-of-rain calls came out almost perfectly calibrated when Murphy and Winkler checked them, and the reason isn't that meteorologists are built differently: they make explicit probabilistic predictions daily and get scored against the outcome every time.
A decision journal runs the same loop on your own life: you state a probability, and the score accumulates as the outcomes come in.
Practice with scoring improves forecasting itself
The Good Judgment Project — the multi-year forecasting tournament behind Tetlock's Superforecasting — tested this at scale. Across the four-year series, Chang and colleagues found that a probabilistic-reasoning module taking under an hour improved Brier scores by 6 to 11% over the control condition. The tournament's broader finding: the forecasters who kept improving were the ones putting numbers on their predictions and revising those numbers as information came in, and then going back over the misses afterwards. Those are the prediction, the check-ins and the resolution entry of a decision journal, under different names.
This is the piece I'd hold most loosely, though. A 2025 reanalysis ran the first two years of the same tournament through an item-response model that controls for which questions forecasters picked, when they answered, and how hard those questions were — all of which varied with the experimental conditions, and none of which the original design controlled. The training and teaming effects Mellers and colleagues had reported shrank, disappeared, or in places reversed. It only covers the first two years while Chang's numbers run across four, so it doesn't settle the later ones, but it's the same tournament and the same design underneath.
What this doesn't prove
The part I'd push back on hardest is transfer. Every forecasting result above comes from tournaments about world events, scored against a public answer; a career move has no public answer and you only get one go at it. That the skill carries from one to the other seems plausible to me, but nobody has tested it. Calibration training also generalizes imperfectly across domains, and journaling has a selection problem in the wild, since the people who keep one already care about their judgment. If someone runs a proper trial of the full practice someday, I'll link it here whichever way it comes out.
There's an endorsement in the mix too. When Michael Mauboussin asked Daniel Kahneman what single thing an investor could do to improve performance, Kahneman's answer was a cheap notebook: write down what you expect to happen and why, so hindsight bias can't later convince you that you knew things you didn't. That's authority rather than data, but it's authority from the person who mapped most of the biases involved.
Where that leaves it
The failure mode is real and well-documented. The measurement it requires is cheap: ten minutes per consequential decision. The mechanisms it relies on have independent support, the forecasting-training one contested, and the one number circulating in its favor is the one thing I'd throw out. On that evidence I keep a journal, with a free template if you'd like to start. Run it for twenty decisions and score yourself; your own calibration curve will convince you more than any study I could have cited here.
More guides in Making better decisions.
Full disclosure: I built Reckon, an iOS app that runs this loop and does the scoring for you — prediction, confidence, check-ins, resolution, the delayed second look — so the measurement stays cheap enough that you actually keep doing it. One-time, no subscription. The template above works on paper; the app is for when you'd rather not tally the buckets by hand.