What Happens to a Prediction After the Future Arrives?

A prediction has a strange life.

Before an event, it is exposed.

The future can still contradict it.

After the event, everything changes. The result is known. Explanations multiply. What once looked uncertain begins to look inevitable. A forecast that was made in public can quietly become a story about what somebody supposedly “knew all along”.

Clyde Street offers a useful defence against that kind of hindsight.

Keith Lyons kept writing before events as well as after them. Medal projections appeared months before an Olympic Games. World Cup models were compared before the tournament. A Grand Final post stated exactly what information he had chosen to use, and what he had deliberately chosen not to know. Then later posts encountered the result.

The interest is not that Keith was always right.

He was not.

The interest is that the record often lets us see the forecast while the future was still open.

That changes the question.

Instead of asking only, “Was the prediction correct?”, we can ask something harder:

What does a prediction owe the future after the future arrives?

Keep the before-state

In January 2012, London was still more than six months away.

Keith published the January projections from USA Today’s Olympic Medal Tracker. The service used an algorithm that monitored athletes’ performances, projected finishes in medal events and accumulated those results into country projections. It was updated monthly.

The model was not Keith’s.

That matters. Clyde Street was carrying an external forecast into view, not claiming ownership of the underlying algorithm.

What the post does give us is a dated state.

On 5 January, this is what one forecasting system was saying about an event that had not happened.

There were more London prediction posts as the Games approached. Different services appeared. Their methods and inputs were not necessarily the same. That makes the archive less tidy, but more useful. It would be wrong to treat every medal projection as one continuous Keith-owned model.

What survives is a sequence of prospective states.

By 13 August, the Games were over. Keith returned to medal prediction through a different service, the Herald Sun’s Virtual Medal Predictor. He placed a prediction made twenty-six days before the Games beside the final London medal table and then looked across Sydney, Athens, Beijing and London.

The preserved text does not give us enough to reconstruct every number in those tables here.

But it gives us something more fundamental: a before and an after that remain distinguishable.

That distinction is easy to undervalue.

A forecast is not simply a statement about the future. It is a statement made from a particular information state. If the earlier state disappears, the later reader loses the ability to tell what was actually expected before the outcome became known.

Clyde Street therefore gives the forecast an afterlife without allowing the result to rewrite its birth certificate.

A forecast can change before the future arrives

Preserving the before-state does not mean freezing every forecast at its earliest version.

Events unfold. Evidence changes. Models update. A person can revise a judgement while the outcome is still unknown.

The 2014 FIFA World Cup makes that distinction unusually clear.

On 30 May, before the tournament, Keith assembled a crowded field of forecasts.

Goldman Sachs used historical results, Elo differences, home advantage and simulation. Andrew Yuan used logistic regression on FIFA match history; David Dormagen used a simulation that could combine rating systems. Opta, Ray Stefani, betting markets and, in a postscript, Infostrada supplied other forms of expectation.

Several routes pointed in the same direction.

Brazil appeared repeatedly among the favourites. Goldman Sachs projected Brazil, Germany, Argentina and Spain into the semi-finals and Brazil over Argentina in the final. Keith noticed what he called an ensemble convergence.

That convergence is historically interesting.

It is not validation.

A collection of models can agree before an event and still share blind spots, assumptions or simply encounter a future that does not cooperate. Agreement tells us something about the expectation landscape. It does not tell us that the event has already been solved.

Six weeks later, on the day of the Final, the information state was different.

Brazil was no longer a possible finalist. Germany’s performance had changed what Keith was considering. Infostrada’s final forecast put Argentina and Germany almost exactly level. Keith looked at experience, Elo ratings, earlier tournament performance and the 2010 meeting between the sides.

Then, before the game, he made his own judgement visible: on balance, he thought it was Germany’s Final.

That is not the same act as the May post.

In May, Keith was curating and comparing models whose ownership remained with their authors. In July, after a month of tournament evidence but before the Final itself, he was making a Keith-owned prospective judgement.

Both belong in the record.

The difference matters because updating before an outcome is not hindsight.

If a forecast changes because new evidence has arrived while the future remains unresolved, the change can be part of responsible prediction. What would be misleading is to erase the earlier state and present the later judgement as though it had always been the view.

Clyde Street lets us see the movement.

The day after the Final, Keith’s tournament record moved on to the 171 goals scored across the competition, with Mario Götze’s 113th-minute goal recorded as the last one.

The future had arrived.

What happened next did not alter what the May forecasts had been. Their value as evidence depends on our ability to keep them in May.

Declare what you chose not to know

A prediction becomes especially revealing when its information boundary is explicit.

Keith’s 2017 AFL Grand Final posts provide one of the clearest examples in Clyde Street.

On the eve of the Final between Adelaide and Richmond, he described exactly how narrow his view of the competition had been.

He had manually recorded quarter-by-quarter scores from the official AFL site.

He had not watched a minute of play.

He did not read the newspaper coverage or listen to football programmes during the season.

He was intentionally asking how far one indicator — the score at the end of each quarter — might take him.

Other forecasts were visible in the post, including bookmaker odds and predictions from Darren O’Shaughnessy and Tony Corke. But Keith’s own judgement rested on his median scoring profiles.

Those profiles suggested an Adelaide advantage, all other things being equal.

Even before the game, the record was not perfectly static. Keith noted that he corrected the data on 30 September and revised the estimated advantage after reviewing it, still before the game.

That is precisely the kind of alteration a prospective record should preserve.

The revision does not weaken the forecast merely because it changed.

It tells us what changed, and when.

Keith also stated conditions that might make Richmond dangerous: a close first quarter, staying with Adelaide in the second, lifting in the third, and then having the legs and tactical nous for the fourth.

He was not building an all-seeing model.

He was putting a deliberately narrow proposition into the open.

That makes the eventual result more informative, not less.

Let the result answer back

Richmond won the Grand Final by forty-eight points.

Two days later, Keith returned to the forecast.

The preserved earlier post still carries the Adelaide expectation.

He began from the discrepancy.

His “all other things being equal” model had indicated an Adelaide win. The game had not behaved that way. Richmond turned an eleven-point first-quarter deficit into a half-time lead, then extended that advantage through the third and fourth quarters.

Some of the conditions Keith had identified as routes into the game for Richmond were visible in the scoring pattern.

This is the point at which the forecast article could easily become an article about learning from failure.

Clyde Street certainly goes there. The result sends Keith to George Lewis and Jonathan Lewis on negative evidence and then into questions about learning environments, improvisation and moving from IF…THEN towards YES…AND.

That is a real Clyde Street route.

It is not the one this article needs to own.

That different question — what changed in the inquiry after the failed expectation — is followed in When the Inquiry Itself Changes.

For the forecast problem, the crucial fact is simpler.

The pre-event claim survives beside the contrary outcome.

We can see what Keith knew. We can see what he excluded. We can see that he corrected the data before play. We can see the expectation. Then we can see what happened.

That sequence protects both sides of the encounter.

It protects the forecast from being caricatured after the event as more confident or comprehensive than it was.

And it protects the outcome from being forced to confirm the model.

A failed forecast does not automatically prove the opposite theory.

A successful forecast would not have proved the whole model either.

The result answers the prediction.

It does not explain itself.

Sometimes the archive is not clean enough

There is a further complication.

A post can look prospective without giving us a clean prospective state.

On 2 October 2017, Clyde Street contains a post titled Points Scored Profiles Prior To #NRLGF 2017.

The title sounds ideal for this article.

The preserved body is not.

It opens by stating that Melbourne Storm had defeated North Queensland Cowboys in the Grand Final. Later in the same preserved text, Keith explains that the scoring profiles had led him to believe Melbourne would win by at least nine points.

Both statements can be true.

But the surviving page is already outcome-aware.

That means we cannot use the preserved version alone as pristine evidence of what the page said before the match.

The title does not rescue us.

Nor does the date by itself.

This is not evidence that Keith rewrote the forecast after the event, and it should not be treated as such. There may have been an earlier state, an update, a publication-timing issue or another route through the material.

The responsible conclusion is narrower.

The snapshot we have is not a clean pre-event source.

If we wanted to make a historical claim about the exact prior forecast, we would need another contemporaneous record.

That small problem exposes something important about prediction archives.

The prediction does not only need to survive.

Its state needs to survive.

When was this version captured?

What did it contain at that point?

Was it updated before or after the outcome?

A prospective title attached to an outcome-aware body is a reminder that chronology itself can require verification.

From prediction to verification

By late 2018, Clyde Street’s forecasting questions had become more explicit.

Keith wrote a post called Forecasting, predicting, classifying and uncertainty after “bumping into” the work of Glenn Brier, Frank Harrell and David Spiegelhalter. The immediate context included an email exchange with Tony Corke about how to share posterior outcomes in relation to prior probability statements from Keith’s Women’s T20 cricket data.

The language changes here.

Brier’s work makes forecast verification and probability statements explicit. Harrell pushes against collapsing prediction into classification. Spiegelhalter’s work brings risk and the communication of uncertainty into view.

These are not Keith’s theories.

They belong to the people who developed them, and the post is clear about the route by which Keith encountered them.

Nor should we use the 2018 post to pretend that every medal projection or football forecast on Clyde Street had secretly been a formal Brier-score exercise all along.

It had not.

What changes is the vocabulary available for asking better questions of forecasts.

How should a probabilistic statement be verified?

What does the user need to know about a forecast’s reliability?

Are we predicting a probability, or merely assigning a class?

How should uncertainty be communicated rather than hidden behind a single confident number?

The earlier Clyde Street archive gives these later questions something to work on.

It already contains pre-event expectations, updates, rival models, narrow information sets, corrected inputs, contrary outcomes and preserved post-event reflections.

The later statistical encounters do not validate the earlier practices.

They make their unresolved problems easier to name.

What a prediction owes the future

Clyde Street does not give us one forecasting method.

It gives us something more useful for this particular question: a long record of forecasts in different states.

Some are Keith’s. Many belong to other forecasters and are curated or compared by him. Some are updated as new evidence arrives. Some are contradicted. Some are followed by reflection. At least one apparently prospective source cannot safely be treated as a clean prior state from the preserved page alone.

That messiness is the point.

A prediction is accountable to more than its final scorecard.

It owes the future a visible past.

We need to know when the forecast was made. We need to know whose forecast it was. We need to know what information it used, and sometimes what information it deliberately excluded. We need to keep revisions on the correct side of the event. We need to distinguish agreement among models from evidence that a model is valid. We need the outcome without allowing the outcome to rewrite the expectation that preceded it.

And if the source itself has changed state, we need enough provenance to say that too.

This is why a correct prediction can still be a poor explanation, and a wrong prediction can still be an intellectually useful object.

The value is not only in whether it hit the target.

It is in whether another reader can reconstruct the distance between what was known then and what is known now.

The future will always have an unfair advantage over the forecast.

It already knows what happened.

A good record gives the forecast one defence: it lets us meet it before the ending.

Question to carry

When the outcome is known, can another person still reconstruct what was predicted, when, by whom, from what evidence, and what changed afterwards?

Publication boundary note

This article is about forecast accountability across source states, not about proving that Keith’s forecasts were accurate or that failure was productive. External forecasting models remain owned by their authors. Later statistical frameworks are not retroactively imposed on earlier Clyde Street work. Pre-event revision is kept distinct from post-event rewriting. A correct outcome does not validate a whole model; a failed outcome does not validate a replacement explanation. Where the preserved source is already outcome-aware, the article says so rather than reconstructing a cleaner prior state than the evidence supports.