The future does not exist yet

The past can be observed.

The present can be experienced.

The future cannot.

That simple fact sits at the heart of prediction.

We often talk about the future as though it were already there, waiting to be discovered:

tomorrow's weather,
next week's electricity demand,
whether a customer will repay a loan,
whether a machine will fail,
whether a patient will recover,
whether a person will click on a recommendation.

But these things have not happened yet.

There is no hidden dataset containing the answer.

There are only possible futures.

Prediction is the process of reasoning about those possibilities.

A forecast is not a fact

Suppose a weather forecast says:

There is a 70% chance of rain tomorrow.

This does not mean that it will rain for 70% of the day.

It does not mean that 70% of the city will get wet.

And it does not mean that the forecaster is 70% certain in some vague sense.

It means that, given the information available and the forecasting model being used, rain is considered more likely than not.

The forecast is therefore not a statement of fact.

It is a statement about uncertainty.

The future remains unresolved.

Possible futures

Imagine that today is Monday.

We are trying to predict electricity demand at 18:00 tomorrow.

There is not one known future value waiting for us.

There are many plausible possibilities.

Demand might be:

40 GW,
42 GW,
45 GW,
48 GW,
50 GW.

Some of those values may be more likely than others.

A forecasting model might produce something like:

Demand	Probability
40 GW	5%
42 GW	15%
45 GW	40%
48 GW	30%
50 GW	10%

The most likely outcome is 45 GW.

But there is still a substantial probability that demand will be higher or lower.

A useful forecast therefore does not necessarily say:

Demand will be 45 GW.

It might instead say:

Demand is most likely to be around 45 GW, but there is meaningful uncertainty around that estimate.

This is a much more accurate description of what prediction actually means.

Forecasts are distributions

A powerful way to represent uncertainty is with a probability distribution.

Instead of predicting one value, we assign different probabilities to different possible outcomes.

Conceptually:

POSSIBLE FUTURE 1 → some probability

POSSIBLE FUTURE 2 → some probability

POSSIBLE FUTURE 3 → some probability

and so on.

Taken together, these possibilities form a distribution over possible futures.

This gives us much more information than a single point prediction.

For example:

Tomorrow's temperature will be 18°C.

is less informative than:

Tomorrow's temperature is expected to be around 18°C, with most plausible outcomes between 15°C and 21°C.

The second statement tells us something about both the expected outcome and the uncertainty around it.

Point predictions hide uncertainty

Machine-learning systems often return a single number.

A house might be predicted to sell for:

€420,000

A journey might be predicted to take:

37 minutes

A person's probability of default might be:

A wind farm might be predicted to generate:

310 MW

Single values are convenient.

But they can create an illusion of certainty.

Imagine two forecasting models.

Model A predicts tomorrow's electricity demand will be 45 GW, with a very narrow range of plausible outcomes.

Model B also predicts 45 GW, but with a very wide range.

Both have the same headline forecast.

But they are not telling us the same thing.

One is relatively confident.

The other is highly uncertain.

For decision-making, that difference can matter enormously.

Decisions depend on uncertainty

Suppose a hospital expects 100 patients tomorrow.

If the likely range is:

98 to 102 patients

then planning for 100 may be perfectly reasonable.

But if the likely range is:

60 to 160 patients

then the same average forecast tells us much less.

The expected value has not changed.

The uncertainty has.

A decision-maker therefore needs to ask two questions:

What do we expect to happen?

and:

How uncertain are we?

This distinction becomes especially important in systems where the cost of being wrong is high.

Expected values are not guaranteed outcomes

Suppose a model predicts that a machine has a 10% chance of failure tomorrow.

We cannot observe a 10% failure.

The machine will either fail or it will not.

The probability describes uncertainty before the outcome occurs.

Likewise, if a coin has a 50% probability of landing heads, a single toss will still produce either heads or tails.

Probability is not describing a partially realised outcome.

It describes our uncertainty across possible outcomes.

This is one reason probabilistic thinking can feel unintuitive.

Real events happen as concrete outcomes.

Forecasts describe the possibilities that existed beforehand.

Uncertainty is not the same as ignorance

There is another important distinction.

Sometimes we know the possible outcomes reasonably well but do not know which one will occur.

That is uncertainty.

Sometimes we do not even know enough to describe the possibilities properly.

That is closer to ignorance.

Consider rolling a fair six-sided die.

We know the possible outcomes:

1, 2, 3, 4, 5 or 6.

We also have a reasonable model of their probabilities.

We are uncertain about the result, but we are not especially ignorant about the process.

Now imagine predicting the economic consequences of a completely new technology twenty years from now.

We may not know:

which industries will exist,
how regulation will change,
what social norms will emerge,
what new technologies will interact with it,
what political events will occur.

Here the problem is deeper.

We may not even know all the relevant future possibilities.

That is not merely uncertainty.

It is uncertainty combined with incomplete knowledge.

Known unknowns and unknown unknowns

A useful distinction is between:

known unknowns

and:

unknown unknowns

A known unknown is something we know is uncertain.

For example:

We know tomorrow's wind speed is uncertain.

We can often build a probability distribution around it.

An unknown unknown is something we have not represented at all.

For example:

A previously unknown failure mode appears in a new battery chemistry.

The model cannot assign a sensible probability to something it does not know exists.

This creates a fundamental limit to forecasting.

Models can only represent uncertainty over possibilities they are capable of representing.

They cannot reliably quantify what lies completely outside their model of the world.

Models create bounded worlds

Every model simplifies reality.

Suppose we build a model of traffic flow.

It might include:

road capacity,
vehicle speed,
traffic signals,
historical congestion,
weather.

But perhaps it does not include:

a major accident,
a political protest,
a burst water main,
a football celebration,
thousands of people suddenly changing route because of another app.

Within the model, predictions may be highly sophisticated.

But the model still describes only a bounded version of reality.

This is why a probability distribution produced by a model should not be interpreted as a perfect description of all possible futures.

It describes possibilities within the assumptions of the model.

Uncertainty comes from different places

Not all uncertainty has the same cause.

Some uncertainty comes from randomness.

Some comes from incomplete measurement.

Some comes from noisy data.

Some comes from imperfect models.

Some comes from human behaviour.

Some comes from future decisions that have not yet been made.

And some comes from events we have not imagined.

This matters because different forms of uncertainty may require different responses.

More data can sometimes reduce uncertainty.

Better sensors can reduce measurement error.

Better models can improve forecasts.

But some uncertainty is inherent in the system itself.

No amount of historical data can make the future completely observable.

The future branches

One useful way to imagine the future is as a branching structure.

From the present state, several things could happen next.

Each of those futures creates another set of possibilities.

Conceptually:

NOW

→ Future A

→ Future B

→ Future C

And from Future A:

→ A1

→ A2

→ A3

From Future B:

→ B1

→ B2

And so on.

As we look further into the future, the number of possible paths can grow enormously.

This is one reason long-term forecasting is generally harder than short-term forecasting.

The further ahead we look, the more future events, decisions and disturbances can intervene.

Prediction horizons matter

Consider predicting where a car will be.

One second into the future may be relatively easy.

If we know its current position, velocity and direction, we can make a useful estimate.

One minute into the future is harder.

The driver may turn, brake or accelerate.

One hour into the future is much harder again.

The driver may stop, change destination, encounter traffic or leave the road entirely.

The same principle applies across many systems.

Weather forecasts usually become less certain further into the future.

Electricity forecasts become harder over longer horizons.

Economic forecasts accumulate more uncertainty over time.

The future becomes less constrained by the present as the prediction horizon grows.

Some systems are more predictable than others

Predictability depends on the nature of the system.

The movement of planets can be predicted with remarkable accuracy.

The behaviour of individual people is much harder.

Weather can often be predicted reasonably well over short horizons but becomes increasingly uncertain further ahead.

Financial markets can respond to information, expectations and the predictions of other participants.

Social systems are particularly challenging because people can change their behaviour in response to forecasts themselves.

Prediction therefore depends not only on how much data we have, but on what kind of system we are trying to predict.

Prediction can influence the future

Suppose a model predicts that a product will sell out tomorrow.

A retailer responds by ordering more stock.

The product does not sell out.

The forecast helped prevent the predicted outcome.

Alternatively, suppose an influential analyst predicts that a bank is in trouble.

Customers react by withdrawing their money.

The prediction may help create the event it predicted.

Predictions can therefore be:

self-defeating,
self-fulfilling,
or largely irrelevant to the thing being predicted.

This becomes especially important when artificial intelligence is embedded inside social and economic systems.

The future is not always independent of the forecast.

Better prediction does not remove uncertainty

Modern AI can dramatically improve our ability to recognise patterns.

It can process more data.

It can identify relationships humans might miss.

It can make predictions faster.

It can update predictions as new observations arrive.

But even an extremely capable AI does not convert the future into a known fact.

A better model may make the probability distribution narrower.

It may improve calibration.

It may reduce error.

But uncertainty remains.

This gives us an important principle:

Intelligence can improve prediction without making the future deterministic.

That distinction becomes crucial when people place too much confidence in automated systems.

Confidence should match evidence

A good forecasting system should not merely be accurate.

Its confidence should also be appropriate.

Imagine a model makes 100 predictions, each labelled:

90% probability

If the model is well calibrated, roughly 90 of those events should occur.

If only 50 occur, the model is overconfident.

If all 100 occur, it may be underconfident.

This idea is called calibration.

A good probabilistic model should know something about what it does not know.

That is often just as important as producing the prediction itself.

We will explore calibration and other ways of evaluating predictions later in the course.

Uncertainty changes how we should act

Imagine two bridges.

An engineer estimates that Bridge A can safely carry between 99 and 101 tonnes.

Bridge B is estimated to carry somewhere between 60 and 140 tonnes.

Both estimates have an expected value near 100 tonnes.

Would you treat them as equally safe for a 95-tonne load?

Probably not.

The uncertainty matters.

This is why decision-making under uncertainty is different from simply choosing the action associated with the average forecast.

Sometimes we need:

safety margins,
reserves,
redundancy,
insurance,
contingency plans,
alternative options.

The more uncertain the future, the more valuable flexibility can become.

AI predicts possibilities from the past

We can now extend the framework from the previous lessons.

The past gives us observations:

PAST → DATA

Those observations help us construct a representation of the present:

DATA → MODEL / STATE

The model allows us to generate possible futures:

MODEL → FORECAST

But the forecast should not be understood as:

this is what will happen

Instead, it should be understood as:

given what we know now, these are the futures we consider plausible.

That distinction lies at the heart of responsible prediction.

The present continually collapses uncertainty

There is something interesting about the passage of time.

At 08:00, many different outcomes may be possible at 09:00.

At 08:30, new information arrives.

Some possibilities become less likely.

Others become more likely.

At 08:59, uncertainty may have narrowed substantially.

Then 09:00 arrives.

One particular outcome occurs.

The uncertain future becomes an observed present.

A moment later, it becomes part of the past.

So our conceptual loop becomes:

PAST → PRESENT STATE → POSSIBLE FUTURES

Then time advances:

POSSIBLE FUTURES → ONE REALISED OUTCOME

And that realised outcome becomes:

NEW DATA

Prediction is therefore an ongoing process.

As the present changes, our forecasts should change with it.

The central idea

The future is not a dataset waiting to be read.

It is a set of possibilities.

Some possibilities may be much more likely than others.

Some uncertainty can be quantified.

Some cannot.

And some possibilities may lie outside our model entirely.

A forecast is not knowledge of the future. It is a structured statement about uncertainty from the perspective of the present.

Artificial intelligence can help us make those statements more accurately.

But it cannot remove the distinction between what has already happened and what has not.

This gives us the next part of our course framework:

PAST → DATA → PRESENT STATE → POSSIBLE FUTURES

In the next lesson, we will examine how prediction creates the bridge between those two sides of time: how information about the past and present can be used to infer what may happen next.