STRATEGY

Months to result

Prediction Autopsy and Model Rebuild

Turn consistent forecast failures into a missing-variable hunt that beats consensus.

forecasting mental models contrarian thinking analytical frameworks economic thinking prediction intellectual honesty

Problem it solves

Analysts and strategists keep repeating the same directional errors because they never interrogate what variable their model is ignoring.

Best for

Traders, investors, strategists, or researchers who rely on forecasts and notice they are consistently wrong in the same direction.

Not ideal for

Operators who need quick tactical decisions and have no historical prediction record to audit against.

Overview

Why this framework exists

Most prediction failures are not random noise—they cluster in a consistent direction, which signals a missing variable in your model. This framework forces you to treat a run of bad forecasts as a diagnostic signal rather than bad luck. You audit the direction of your errors, hypothesise the overlooked variable causing the bias, rebuild your model around that variable, and then validate the updated model by making multiple simultaneous non-consensus predictions. Validation requires stacking several orthogonal calls at once so that a correct outcome cannot be attributed to luck. The process mirrors the scientific method applied to real-world forecasting where controlled experiments are impossible.

Core principles

6 total

Consistent directional errors signal a structural gap in the model, not random bad luck.
The missing variable is usually something the mainstream consensus has ideological or financial reasons to ignore.
In fields where controlled experiments are impossible, track records of non-consensus predictions are the only legitimate credibility signal.
Making multiple simultaneous orthogonal predictions removes the luck explanation from a correct call.
Intellectual honesty requires auditing your own forecasts with the same rigour you apply to competitors' forecasts.
Gatekeeping by credential rather than track record protects wrong incumbents, not truth.

Steps

6 steps

Audit your prediction record for directional bias
Compile every material forecast you made over the last one to three years. Label each right or wrong, then look for a pattern: are you consistently too optimistic about growth, too bearish on inflation, or systematically off in one sector? Random errors spread evenly; a structural gap clusters.
Pro tipKeep a dated prediction log going forward—even a simple spreadsheet—so your next audit is quantitative rather than from memory.
WarningAvoid cherry-picking predictions you remember. Include the embarrassing misses or the audit is useless.
Identify the consistent direction of your errors
Once the audit is done, ask: in what direction am I usually wrong? This is your diagnostic signal. A persistent directional error is evidence that your model systematically underweights or ignores a real-world force.
Pro tipCompare your error direction to the consensus. If you and the consensus share the same directional bias, you are almost certainly missing the same variable.
Hypothesise the missing variable
Ask: 'What single variable, if I had weighted it more heavily, would have corrected my directional bias?' Look at heterodox thinkers, practitioners in adjacent fields, and historical analogues for candidate variables. Gary Stevenson identified wealth distribution and inequality as the variable mainstream models systematically excluded.
Pro tipThe missing variable is often one the dominant community has structural incentives to ignore—check who benefits from the current model being correct.
WarningBeware of retrofitting: the variable must explain the direction of your past errors, not just be consistent with outcomes.
Rebuild the model incorporating the new variable
Revise your analytical framework to give explicit weight to the candidate variable. Write down the mechanism: how does this variable flow through to the outcomes you forecast? A mechanism you can articulate is more robust than a correlation you noticed.
Pro tipKeep the old model running in parallel for at least one cycle so you can see the delta in predictions.
Stack multiple simultaneous non-consensus predictions
Before outcomes are known, publicly commit to several orthogonal predictions derived from the updated model—e.g., rising inflation AND rising house prices AND rising gold simultaneously. Stacking removes the luck explanation: getting one contrarian call right can be chance; getting four interconnected calls right at once is a model signal.
Pro tipTimestamp and publish predictions before consensus has shifted—a prediction made after events start to unfold does not count as non-consensus.
WarningDo not over-specify timelines if the mechanism is structural rather than cyclical; attach a range rather than a single date.
Track outcomes and iterate
Review each prediction against outcomes every quarter. If the new model reduces directional bias substantially, treat it as validated. If bias persists or reverses, return to Step 3. Model revision is iterative, not a one-time fix.
Pro tipDocument the review publicly or with an accountability partner—the discipline of publishing forces honest scoring.
WarningA model can be right for structural reasons in one regime and wrong in another. Validate continuously rather than declaring permanent victory.

Checklist

Saved in your browser

List every major prediction you made over the past 12–24 months and mark each as correct, wrong, or incomplete.
Identify the directional bias: are your errors consistently too optimistic, too pessimistic, or skewed in one domain?
Ask: 'What variable, if I added it to my model, would have flipped the wrong calls?'
Research whether any minority or heterodox thinkers have been emphasising that variable—read their primary sources.
Rebuild your model explicitly incorporating the candidate variable.
Make at least three simultaneous, specific, non-consensus predictions based on the updated model and publish them with a timestamp.
Track outcomes quarterly and document whether the new model reduces directional bias.
Repeat the audit annually—models can become stale as the world changes.

Examples

3 cases

Gary Stevenson's Post-2008 Interest Rate Model

Working as an interest-rate trader from 2008 to 2011, Gary watched mainstream economists predict aggressive recovery—and higher rates—every single year. He audited the consensus track record, identified wealth inequality as the ignored variable suppressing demand, and rebuilt his model around it. He then bet on rates staying low for a structurally long period, generating millions while consensus forecasters were wrong for 13 consecutive years.

OutcomeCorrect non-consensus prediction of persistently low interest rates across 2008–2020, later validated by a second cycle of correct macro calls around the COVID inflation episode.

COVID-Era Multi-Prediction Stack

At the onset of COVID in 2020, Gary published a cluster of simultaneous non-consensus predictions: a cost-of-living crisis, paired with rising house prices, rising stock prices, rising gold, and high inflation. Each individual call was contrarian; all five together were near-impossible to attribute to luck. The mechanism was the same missing variable: deficit spending concentrated gains at the top while suppressing real wages at the bottom.

OutcomeAll five predictions materialised within 18–24 months, providing a strong out-of-sample validation of the updated model.

Plausible Application: Startup Demand Forecasting

A SaaS founder notices her churn models consistently underpredict churn in enterprise accounts. She audits 18 months of predictions, finds they all overestimated retention, and identifies the missing variable: procurement-cycle budget freezes that her model treated as one-off events. She rebuilds the model with a budget-freeze index as an input, then makes three simultaneous predictions about Q3 churn, expansion revenue, and NPS before the quarter opens.

OutcomeNew model reduces churn forecast error by 40% in the first cycle; the multi-metric stack rules out coincidence and gives the board confidence in the revised go-to-market plan.

Common mistakes

3 traps

Auditing only memorable predictions

Analysts naturally recall their correct calls vividly and forget quiet misses. If the audit pool is biased toward remembered predictions, the directional signal is contaminated and the missing variable hunt starts from a false baseline.

Retrofitting the variable after outcomes are known

Identifying a 'missing variable' after results are public is not model correction—it is post-hoc rationalisation. The variable must be hypothesised before the next round of predictions are tested, or the validation step is meaningless.

Making a single big prediction instead of stacking

One contrarian call, even if correct, can always be dismissed as luck. The stacking step—multiple simultaneous orthogonal predictions—is what distinguishes a genuine model upgrade from a coin-flip win. Skipping it leaves the credibility question open.

Origin story

How this framework came to be

Gary Stevenson, LSE economics graduate and interest-rate trader, noticed from 2008 to 2011 that mainstream economists predicted recovery every single year and were wrong every single year. He identified wealth inequality as the missing variable, rebuilt his model, and went on to correctly predict the post-COVID cost-of-living crisis, rising house prices, rising stock prices, rising gold, and elevated inflation—none of which were consensus calls.

Source

Traced to primary

Source · VIDEO

Have I been wrong on the economy the whole time? — Garys Economics

Garys Economics · 2026

Open source →

Related frameworks

Browse all Strategy →