Prediction as a bridge through time

The past has happened.

The future has not.

Yet almost every meaningful decision we make in the present depends on what we think will happen next.

Should we take an umbrella?

How much food should a supermarket order?

How many nurses should a hospital schedule tomorrow?

How much electricity will people need this evening?

Will a machine fail before its next maintenance window?

Is another vehicle about to cross our path?

In each case, we face the same fundamental problem:

We must make decisions now about events that have not happened yet.

Prediction is the bridge between those two sides of time.

We use observations from the past, together with our understanding of present conditions, to infer possible future states.

From observations to predictions

Suppose we measure the outside temperature every hour:

Time	Temperature
08:00	10°C
09:00	11°C
10:00	13°C
11:00	15°C
12:00	16°C

At 12:00, all of these measurements are available to us.

We now want to know:

What might the temperature be at 13:00?

We cannot measure the 13:00 temperature yet.

It does not exist as an observation.

Instead, we infer it.

Perhaps we notice that the temperature has generally been increasing through the morning.

A very simple model might therefore predict:

13:00 → approximately 17°C

We have used a pattern in previous observations to estimate something that has not yet happened.

This is prediction in its simplest form.

The basic structure of prediction

We can represent the process as:

PAST OBSERVATIONS + PRESENT CONDITIONS → MODEL → POSSIBLE FUTURE

The model provides a relationship between what we know and what we want to infer.

For example:

historical weather + current weather → weather model → tomorrow's weather

or:

past electricity demand + current conditions → forecasting model → future demand

or:

previous words + current context → language model → probable next words

The models may be very different.

But the underlying structure is remarkably similar.

Prediction takes information available now and uses it to reason about something that is not yet observable.

Prediction always starts from the present

Every prediction is made from a particular point in time.

Suppose we want to predict electricity demand tomorrow at 18:00.

If we make the prediction today at 09:00, we have one set of information.

If we update the prediction tomorrow at 12:00, we have much more recent information.

If we update it again at 17:55, we know even more.

The event being predicted has not changed:

electricity demand at 18:00

But our knowledge of the system has.

New observations have arrived.

Weather forecasts may have changed.

Demand earlier in the day is now known.

Unexpected events may have occurred.

Our prediction should therefore change as well.

This gives us an important principle:

A forecast is always conditional on the information available when the forecast is made.

Predictions are not permanent facts.

They are estimates made from a particular informational position in time.

The present gives us our starting point

Prediction does not usually depend on every moment in the past equally.

Often, what matters most is the condition of the system now.

Imagine trying to predict where a moving car will be five seconds from now.

Knowing where the car was three years ago is unlikely to help.

Knowing its:

current position,
current speed,
current direction,
current acceleration

is much more useful.

Together, these variables describe important features of the car's current condition.

We can call this its state.

Conceptually:

CURRENT STATE → MODEL → FUTURE STATE

This idea is extremely powerful.

Rather than thinking about prediction as simply extrapolating historical data, we can think of it as asking:

Given where the system appears to be now, and what we know about how it behaves, where might it go next?

What is a state?

A state is a representation of the information about a system that matters for what happens next.

Consider a battery.

Its state might include:

state of charge,
temperature,
current power,
degradation,
operating limits.

If we know those things, we can begin to reason about what the battery might be capable of doing next.

For a vehicle, useful state information might include:

position,
velocity,
direction,
acceleration,
battery level.

For a hospital:

occupied beds,
available beds,
staffing levels,
patients waiting,
expected admissions.

For an electricity system:

current demand,
generator output,
storage levels,
network flows,
weather conditions.

Different systems require different state descriptions.

The central question is always:

What do we need to know about the system now in order to reason about what may happen next?

We will return to this idea later when we explore state vectors, dynamical systems, control, reinforcement learning and robotics.

The past teaches us how things change

Knowing the present state is useful.

But we also need some understanding of how the system tends to evolve.

That understanding often comes from observing transitions in the past.

We repeatedly observe:

STATE NOW → STATE LATER

Over many observations, relationships begin to appear.

For example:

higher outdoor temperature → higher cooling demand

or:

higher wind speed → greater wind generation

or:

particular vibration pattern → increased probability of machine failure

or:

particular sequence of words → increased probability of a particular next word

The historical data therefore gives us examples of change.

We see:

what conditions existed

and then:

what happened next

A predictive model attempts to learn useful relationships between the two.

Prediction is pattern carried forward

Imagine a supermarket records bread sales every day.

It discovers that demand tends to increase sharply between 17:00 and 19:00 on weekdays.

That is a pattern in the past.

When 16:00 arrives today, the supermarket can use that pattern to anticipate increased demand over the next few hours.

The reasoning is approximately:

under these conditions, this pattern has happened before

therefore:

something similar may happen again

Prediction carries learned structure from observed situations into unobserved ones.

This is one of the fundamental ideas behind machine learning.

A useful model does more than remember historical examples.

It learns relationships that can be applied to situations it has not seen before.

This ability is called generalisation.

The future does not have to repeat the past

There is an obvious problem.

Tomorrow is never exactly the same as yesterday.

Electricity demand tomorrow may resemble electricity demand today, but:

the weather may change,
it may be a holiday,
electricity prices may be different,
a major event may be taking place,
a factory may be closed,
millions of individual decisions may differ.

A good prediction therefore cannot simply ask:

What happened last time?

It needs to ask something closer to:

Given what happened in comparable situations, and given what we know about the situation now, what outcomes are plausible next?

That distinction is important.

Prediction is not simply repetition.

It is inference.

Present conditions provide context

Suppose a shop normally sells 100 umbrellas per day.

A model based only on average historical sales might therefore predict:

100 umbrellas tomorrow

But suppose tomorrow's weather forecast predicts extremely heavy rain.

That new information changes the situation.

A better prediction might be:

250 umbrellas tomorrow

The historical pattern has not disappeared.

But present conditions have changed how we interpret it.

So a more useful representation is:

HISTORY + PRESENT CONTEXT → PREDICTION

Context might include:

weather,
location,
time of day,
day of week,
season,
price,
recent behaviour,
current system state,
external events.

Modern AI systems can incorporate enormous amounts of context.

But the principle remains the same:

The future is inferred from the past through the lens of the present.

Time itself contains information

Time is often an important predictor.

Human activity follows strong temporal patterns.

Electricity demand varies by:

hour of day,
weekday versus weekend,
season,
holidays.

Transport demand changes around:

morning commuting,
evening commuting,
school schedules,
sporting events.

Retail demand changes around:

weekends,
paydays,
Christmas,
seasonal weather.

A measurement therefore does not only have a value.

It also has a time.

Knowing that electricity demand was 35 GW is useful.

Knowing that it was 35 GW at 03:00 on a Sunday tells us something more.

Time provides context.

Sequences contain information

Sometimes a single observation tells us very little.

The sequence leading up to it tells us much more.

Imagine the temperature of a machine is currently 80°C.

Is that dangerous?

We might not know.

Now consider two sequences.

Machine A

60 → 65 → 70 → 75 → 80°C

Machine B

83 → 82 → 81 → 80 → 80°C

Both machines are currently at exactly 80°C.

But their histories are very different.

Machine A is heating rapidly.

Machine B is cooling.

The present value is identical.

The trajectory is not.

This tells us something important:

To predict the future, we may need to know not only where a system is, but how it arrived there.

This is why time-series models, recurrent models, state estimation and memory can be valuable.

Speed and direction are information about the future

Consider a moving object.

Its current position tells us where it is.

Its velocity tells us something about where it is going.

Suppose two cars are currently at exactly the same point.

One is travelling north at 50 km/h.

The other is travelling south at 50 km/h.

Their present positions are identical.

Their likely future positions are very different.

The state therefore often contains information about change, not simply position.

This idea appears throughout science and engineering.

Knowing both where something is and how it is changing provides a much stronger basis for prediction.

Recent information may matter more

Not every observation from the past should necessarily have equal influence.

Suppose we want to predict the popularity of a restaurant.

Reviews from last week may be more relevant than reviews from ten years ago.

Suppose we are monitoring an industrial machine.

A vibration measurement from five seconds ago may matter more than one recorded when the machine was manufactured.

As systems change, older information can become less representative of the present.

This creates an important tension.

More historical data gives us more examples.

But older data may describe a world that no longer exists.

The question is not simply:

How much data do we have?

It is also:

How relevant is that data to the system as it exists now?

When the world changes

Suppose we build a model of commuting behaviour using data collected over many years.

Then working patterns change.

Remote and hybrid working become widespread.

The historical relationships may no longer describe the present very well.

The model may have accurately learned the past.

But the future is being generated by a different system.

This is sometimes called distribution shift.

The statistical environment changes.

That gives us another fundamental principle:

Learning from the past works best when the relationships learned from the past remain relevant to the future.

When those relationships change, predictions can fail.

Models therefore need to be updated as the world changes.

Prediction horizons matter

There is a major difference between predicting:

one second ahead

and:

twenty years ahead.

Imagine predicting where a car will be.

One second ahead may be relatively easy.

Its current position, velocity and direction provide strong information.

Ten seconds ahead is harder.

The driver could accelerate, brake or turn.

Ten minutes ahead is harder again.

The driver could take many possible routes.

Ten hours ahead, the car could be parked hundreds of kilometres away.

The further into the future we predict, the more opportunities there are for:

decisions,
disturbances,
unexpected events,
interactions,
new information

to change the trajectory.

In general:

Uncertainty tends to grow with the prediction horizon.

Different decisions require different futures

There is no single useful forecasting horizon.

A robot may need to predict milliseconds ahead.

An autonomous vehicle may reason seconds ahead.

An electricity system might forecast:

seconds ahead for control,
minutes ahead for balancing,
hours ahead for scheduling,
days ahead for planning,
years ahead for infrastructure investment.

A business might predict:

sales this afternoon,
demand next month,
market size in five years.

Different futures support different decisions.

This means prediction should always be connected to a question:

What decision are we trying to make, and how far ahead do we need to see?

Models can predict trajectories

Sometimes we care about more than one future point.

We want to predict a trajectory.

Suppose a robot sees a pedestrian crossing a road.

It does not only care where the pedestrian will be one second from now.

It might estimate:

NOW → 1 second → 2 seconds → 3 seconds → 4 seconds

This gives the robot a possible path through time.

Weather forecasting does something similar.

Electricity forecasting does something similar.

Economic forecasts often do the same.

A predicted future can therefore be represented as a sequence of future states:

x(t+1), x(t+2), x(t+3), ...

The further along the trajectory we move, the greater the uncertainty often becomes.

There may be many trajectories

The previous lesson introduced an important idea:

the future is not necessarily one path.

Suppose an autonomous vehicle observes another car approaching a junction.

That car might:

turn left,
continue straight,
turn right,
stop.

A good system should not necessarily choose one outcome and treat it as certain.

It may instead consider multiple possible trajectories.

Conceptually:

CURRENT STATE

→ FUTURE A

→ FUTURE B

→ FUTURE C

with different probabilities attached to each.

Prediction is therefore not always about finding the future.

It may be about identifying a useful set of possible futures.

Prediction through space as well as time

Prediction is not only about when something will happen.

It is often also about where.

A weather forecast tells us:

what conditions may occur, where, and when

A transport model predicts:

where vehicles and passengers may be at particular times

A power-system model predicts:

where electricity will be generated and consumed

A robot predicts:

where surrounding objects may move

Many prediction problems are therefore really spatiotemporal problems.

They concern space and time together.

We are not merely asking:

What happens next?

We are asking:

What happens next, where?

We will explore the importance of space more directly in the next lesson.

Prediction can include our own actions

There is another important complication.

Sometimes the future depends on what we decide to do.

Suppose a robot is moving towards an obstacle.

If it continues straight ahead, one future trajectory results.

If it turns left, another future results.

If it stops, another future results.

So the prediction problem becomes:

CURRENT STATE + POSSIBLE ACTION → POSSIBLE FUTURE STATE

This is a major step.

We are no longer simply predicting what the world will do.

We are predicting how the world might evolve under different decisions.

This relationship lies at the heart of planning, optimisation, control and reinforcement learning.

Prediction can become simulation

Suppose we ask:

What happens if we do nothing?

Then:

What happens if we take Action A?

Then:

What happens if we take Action B?

A model can generate possible futures under each option.

Conceptually:

CURRENT STATE

→ Action A → Future A

→ Action B → Future B

→ Action C → Future C

Now prediction is helping us compare decisions.

The model becomes a kind of simulator.

This is one of the most powerful uses of models:

They allow us to explore possible futures before committing to an action in the real world.

Forecasting and planning are different

It is useful to distinguish between two questions.

Forecasting

What is likely to happen?

Planning

What should we do?

Forecasting estimates future states.

Planning evaluates possible actions.

A prediction might tell us:

Electricity demand is likely to reach 50 GW this evening.

Planning asks:

Which generators, batteries, interconnectors or flexible loads should we use to serve that demand?

A model might predict:

Traffic on this road will become congested.

Planning asks:

Should vehicles be routed somewhere else?

This distinction will become central later in the course.

Prediction describes possibilities. Decisions choose among actions.

AI as a prediction machine

One way of understanding much of modern artificial intelligence is as increasingly sophisticated prediction.

A computer-vision system might predict:

Which object produced these pixels?

A speech-recognition system might predict:

Which words produced this sound?

A recommendation system might predict:

What is this person likely to engage with?

A forecasting model might predict:

What value is likely to occur next?

A large language model repeatedly asks something like:

Given everything in the context so far, what token is likely to come next?

These applications look very different.

But at an abstract level they all use information that is available to infer information that is unknown.

Prediction does not require seeing the future

This is worth emphasising.

AI does not somehow obtain information from the future.

It learns relationships from observations that already exist.

It then applies those relationships to current conditions.

The process is:

PAST EXAMPLES

↓

LEARN RELATIONSHIPS

↓

OBSERVE PRESENT CONDITIONS

↓

INFER POSSIBLE FUTURE OR UNKNOWN OUTCOME

This can produce remarkably powerful behaviour.

But the logical structure remains grounded in the past and present.

The bridge is never perfect

Prediction connects the known world to the unknown future.

But the bridge is never perfect.

Errors can arise because:

the data is incomplete,
measurements contain noise,
important variables are missing,
the model is imperfect,
the world has changed,
human behaviour is difficult to predict,
unexpected events occur.

This is why the previous lesson's discussion of uncertainty matters.

A model does not eliminate uncertainty simply because it produces a number.

A useful prediction should be understood in the context of what the model knows, what it does not know, and how far into the future it is trying to see.

Prediction gets updated

The bridge between present and future is continuously rebuilt.

Suppose at 06:00 we forecast electricity demand for 18:00.

At 09:00, new information arrives.

We forecast again.

At 12:00, we update again.

At 17:00, we update again.

Each new observation moves the boundary between past and future.

What was previously predicted becomes observed.

The current state changes.

The forecast changes.

This creates a rolling process:

OBSERVE → UPDATE STATE → PREDICT → OBSERVE → UPDATE STATE → PREDICT → ...

Intelligent systems can perform this process continuously.

Prediction and feedback

Once predictions influence decisions, something deeper happens.

Consider a navigation system.

It predicts congestion.

It sends drivers along another road.

Those drivers change the traffic conditions.

The next prediction now begins from a different state.

The cycle becomes:

OBSERVE → PREDICT → DECIDE → ACT → CHANGE THE WORLD → OBSERVE AGAIN

The prediction is no longer detached from the system.

It is part of a feedback loop.

This will become increasingly important as we move from passive forecasting systems towards AI agents and autonomous machines.

The past, present and future together

We can now combine the ideas from the first four lessons.

The past gives us evidence:

PAST → OBSERVATIONS

Those observations become data:

OBSERVATIONS → DATA

Data and present conditions help us construct an understanding of where the system is now:

DATA → PRESENT STATE

Models describe relationships and change:

PRESENT STATE + MODEL → POSSIBLE FUTURES

Predictions allow us to reason across the boundary between what has happened and what has not:

PAST + PRESENT → PREDICTION → FUTURE

Prediction is therefore a bridge through time.

But eventually predictions become useful because someone does something with them.

So the wider loop becomes:

PAST → DATA → STATE → MODEL → PREDICTION → DECISION → ACTION → NEW STATE

The new state generates new observations.

And the process starts again.

The central idea

Prediction is not magic.

It is inference across time.

We observe what has happened.

We estimate where the system is now.

We learn relationships describing how things tend to change.

And we use those relationships to reason about what might happen next.

The past provides examples. The present provides the starting point. The model provides the relationship. Prediction provides possible futures.

Artificial intelligence dramatically increases our ability to perform this process across vast datasets and extremely complex systems.

But every prediction still begins on the same side of the boundary:

with information that already exists.

In the next lesson, we will add another dimension.

Events do not only happen in time.

They happen somewhere.

And just as losing information about time can damage a prediction, losing information about space can hide some of the most important structure in a system.