What is prediction?

Prediction is the attempt to say something useful about an outcome that has not yet been observed.

At its simplest:

Prediction uses what we know to estimate what we do not yet know.

That unknown outcome might lie in the future:

tomorrow's temperature,
next week's electricity demand,
whether a machine will fail,
how long a journey will take.

But prediction can also concern something that already exists but has not yet been observed:

whether an image contains a tumour,
whether a transaction is fraudulent,
what word was obscured in a document,
what object is behind another object.

So prediction is not exclusively about the future.

It is about inference beyond the information directly available to us.

Prediction connects data to uncertainty

In the previous modules, we developed this chain:

REALITY

↓

MEASUREMENT

↓

DATA

↓

MODEL

↓

UNCERTAINTY ABOUT WHAT WE DO NOT KNOW

Prediction is what happens when we use the model to move from what has been observed to something that has not.

Conceptually:

KNOWN INFORMATION

↓

MODEL

↓

ESTIMATE OF UNKNOWN OUTCOME

That estimate may be:

one number,
one class,
a probability,
an entire distribution.

A simple prediction

Suppose the temperature at:

12:00 = 18°C

13:00 = 19°C

14:00 = 20°C

We might predict:

15:00 ≈ 21°C

We have recognised a pattern and extended it forward.

That is prediction.

But immediately we should ask:

How confident are we?

Perhaps clouds arrive at 14:30.

Perhaps the previous trend was temporary.

A prediction is therefore not a fact about the future.

It is an inference based on available evidence and assumptions.

Prediction is not prophecy

This distinction is fundamental.

A prediction does not reveal a future that already exists waiting to be discovered.

It says:

Given what we currently know and the model we are using, this is what we expect.

So:

PREDICTION ≠ CERTAINTY

A good prediction can be wrong.

A bad prediction can occasionally be right.

Prediction quality must be judged over repeated cases, not only by one realised outcome.

The future does not have to follow the prediction

Suppose a weather model predicts:

70% probability of rain.

It does not rain.

That does not automatically make the forecast bad.

Likewise, a model predicts:

10% probability of failure.

The machine fails.

That does not automatically make the model bad.

Probabilistic prediction concerns distributions over outcomes.

One future is eventually realised.

Prediction begins with a target

Before we can predict, we need to define:

What exactly are we trying to predict?

This is sometimes called the target, output, or response variable.

Examples include:

Y = tomorrow's electricity demand

Y = house sale price

Y = whether an email is spam

Y = next word in a sentence

Y = pedestrian position two seconds from now

If the target is poorly defined, the prediction problem is poorly defined.

Inputs provide evidence

The information used to make the prediction is usually called the:

inputs,
predictors,
features.

For example, to predict electricity demand we might use:

previous demand,
temperature,
hour of day,
weekday,
season.

Conceptually:

INPUTS X

↓

MODEL

↓

OUTPUT Y

This is one of the central structures of machine learning.

Prediction as a function

A simple mathematical representation is:

ŷ = f(x)

where:

x represents the inputs,
f is the model,
ŷ is the prediction.

The symbol:

is often read as:

y-hat

and means:

our predicted value of y.

The true outcome may later be observed as:

Then we can compare:

prediction ŷ

with:

reality y.

The prediction error

Once the real outcome is known, we can calculate an error.

For a numerical prediction:

error = actual - predicted

Suppose:

predicted demand = 40 GW

and:

actual demand = 42 GW.

The error is:

2 GW.

But how we score that error depends on the application.

Later we will study:

absolute error,
squared error,
classification errors,
probabilistic scoring.

Prediction creates a testable claim

This is one reason prediction is powerful.

A model says:

Given X, I expect Y.

Then reality happens.

We can compare:

MODEL

with:

OUTCOME.

This creates feedback.

The model can improve.

The prediction loop

Conceptually:

PAST DATA

↓

LEARN MODEL

↓

NEW INPUT

↓

PREDICT

↓

REAL OUTCOME

↓

COMPARE

↓

LEARN

Prediction is therefore not just output generation.

It can be part of a learning loop.

Prediction from examples

Suppose we have historical house sales.

For each house we know:

size,
location,
age,
sale price.

A machine-learning model tries to learn the relationship between:

house characteristics

and:

sale price.

Then a new house arrives.

Its sale price has not yet been observed.

The model predicts it.

This is supervised learning.

We provide examples where both:

inputs

and:

outputs

are known.

The system learns how to map between them.

Prediction means generalising beyond the examples

A model that simply memorises every training example is not necessarily useful.

Suppose a house has:

exactly the same characteristics as one previously sold.

Memorisation might work.

But usually new cases differ.

The model needs to infer:

Houses with similar characteristics tend to have similar prices.

Then apply that pattern to a case it has never seen.

This is generalisation.

Generalisation is the real challenge

A model may fit historical data perfectly.

That is not enough.

We want it to perform well on:

new data.

The real machine-learning problem is therefore:

Learn patterns from the past that remain useful beyond the past.

That is much harder than memorisation.

Prediction requires assumptions about continuity

Suppose historical data shows:

larger houses tend to cost more.

We use that relationship for a new house.

Implicitly, we assume:

Something about the historical relationship will continue.

Every prediction contains some assumption of continuity between:

observed world

and:

unobserved world.

If that relationship changes, the prediction may fail.

The future must resemble the past enough

Machine learning works because the future is neither:

completely identical to the past

nor:

completely unrelated to it.

If everything repeated exactly, prediction would be trivial.

If nothing persisted, prediction would be impossible.

Prediction is useful because the world contains:

patterns with some persistence.

Prediction is pattern recognition

Suppose we observe:

temperature

and:

ice cream sales.

A model may learn that hotter days tend to have higher sales.

Then given tomorrow's forecast temperature, it predicts sales.

The model has recognised a relationship.

This is the essence of many machine-learning systems:

find statistical patterns

then:

use them to estimate unknown outcomes.

Patterns can be simple

Some predictive relationships are straightforward.

For example:

distance travelled = speed × time.

If we know:

current position,
speed,
time,

we can predict future position using a physical model.

This is prediction too.

Not all prediction requires machine learning.

Physics can make predictions

Suppose we know:

position,
velocity,
acceleration.

Equations of motion can predict where an object will be.

This is a mechanistic model.

It is based on an explicit description of how the system behaves.

Machine learning often takes a different approach.

It learns patterns from data.

Data-driven prediction

Suppose we cannot easily write an equation describing customer behaviour.

Instead we have millions of historical observations.

A machine-learning model may learn:

inputs → outcomes

without requiring a complete mechanistic theory of why each relationship exists.

This is data-driven prediction.

Mechanistic and data-driven models can work together

The distinction is not absolute.

A weather model might combine:

physical equations,
measured data,
learned corrections.

A robot might combine:

mechanical dynamics,
neural-network perception.

A power-system model might combine:

electrical network equations,
statistical demand forecasts.

Intelligent systems often combine several forms of modelling.

Prediction does not require causation

Suppose umbrella use strongly predicts rain.

If our only task is:

Predict whether it is raining,

umbrella observations may be useful.

We do not need to claim:

umbrellas cause rain.

Prediction asks:

What information helps estimate the target?

Causal reasoning asks:

What happens if we intervene?

These are different problems.

Predictive correlations can be useful

Suppose an AI predicts machine failure using:

vibration,
temperature.

Perhaps it does not fully understand the physical mechanism.

If the relationship is stable, the model may still make excellent predictions.

Prediction can therefore succeed without full causal understanding.

But correlation can become fragile

Suppose a model predicts wolves from photographs.

In the training data:

wolves usually appear in snow,
dogs usually do not.

The model may learn:

snow → wolf.

This can produce excellent historical accuracy.

But it may fail on:

dog in snow.

The model learned a predictive shortcut.

It did not learn the intended concept.

Prediction quality depends on deployment context

A model should not be judged only by:

How well did it predict the training data?

The important question is:

How well does it predict the situations where it will actually be used?

This may involve:

new people,
new locations,
future time periods.

Prediction is always relative to a deployment environment.

The training world and deployment world

Conceptually:

TRAINING WORLD

↓

DATA

↓

MODEL

↓

DEPLOYMENT WORLD

If:

deployment ≈ training

generalisation may be good.

If:

deployment differs greatly

performance can deteriorate.

This is distribution shift.

Prediction across time

Forecasting is a special form of prediction where the target lies in the future.

For example:

Y(t+1)

is predicted using information available at:

time t.

Conceptually:

PAST + PRESENT

↓

MODEL

↓

FUTURE

This is where our course's original timeline becomes explicit.

Forecast horizon matters

Prediction one second ahead is different from prediction:

one hour ahead,
one year ahead,
fifty years ahead.

As the horizon increases:

more disturbances can occur,
more decisions can intervene,
more structural change is possible.

Uncertainty often grows.

Every forecast should therefore specify:

Prediction for when?

Prediction across space

Prediction can also involve locations we have not directly observed.

Suppose weather stations provide temperatures at several points.

We want temperature:

between the stations.

This is spatial prediction.

Or perhaps we predict air pollution across an entire city from a limited sensor network.

The question becomes:

Given what happened here, what is likely happening there?

Prediction across time and space

Many important AI systems operate across both.

For example:

Where will this storm be six hours from now?

The target contains:

location

and:

time.

Likewise:

Where will this pedestrian be two seconds from now?

Prediction often means estimating future state in a spatiotemporal system.

A prediction can be a number

Suppose we predict:

tomorrow's temperature = 18.4°C.

This is a regression problem.

The output is numerical.

Other examples include:

demand,
price,
journey time,
sales.

Regression will be introduced later in this module.

A prediction can be a category

Suppose we predict:

spam

or:

not spam.

This is a classification problem.

Other examples include:

disease / no disease,
cat / dog / fox,
fraud / legitimate.

The output belongs to a category.

A prediction can be a probability

Instead of saying:

fraud

the model may say:

P(Fraud) = 0.82.

This is richer.

The final service can decide what to do with the probability.

For example:

below 5% → approve

5–50% → additional checks

above 50% → human review

The prediction and decision remain separate.

A prediction can be an entire distribution

Suppose future demand is uncertain.

A model might output:

Expected demand = 40 GW

plus:

90% prediction interval = 35–47 GW.

Even better, it may provide the full probability distribution.

This tells us much more than one number.

Predictive distributions enable better decisions

Suppose two forecasts have the same mean.

One is highly certain.

One is extremely uncertain.

If we have only the point prediction, they look identical.

With distributions, the downstream system can behave differently.

For example:

low uncertainty → act normally

high uncertainty → maintain extra reserve.

Prediction is information for decision-making

Suppose an AI predicts:

80% probability this patient benefits from treatment.

That prediction may be excellent.

But suppose:

50 patients

could benefit

and:

10 treatment slots

exist.

The prediction does not solve the problem.

Now we need to decide:

Who receives the scarce treatment?

That is an allocation problem.

Prediction is not decision-making

This distinction will become central later.

Prediction asks:

What is likely to happen?

Decision-making asks:

What should we do?

A prediction can inform a decision.

It does not determine it.

Predictions do not contain objectives

Suppose a model predicts:

95% probability of customer cancellation.

Should the company offer a discount?

That depends on:

value of retaining the customer,
cost of discount,
fairness,
business strategy.

The probability does not contain the objective.

Predictions do not contain constraints

Suppose a model predicts:

all 100 patients would benefit from treatment.

But there are:

20 doses.

The model does not create the missing doses.

Constraints exist outside the prediction.

This is why:

prediction + objectives + constraints

leads toward:

optimisation and service design.

Predictions can change behaviour

Prediction can also affect the thing being predicted.

Suppose a traffic system predicts:

Road A will be congested.

Drivers see the forecast and choose Road B.

Now:

Road A becomes less congested

and:

Road B becomes more congested.

The prediction changed the future.

Reflexive prediction

This occurs whenever people or machines respond to a forecast.

Examples include:

electricity prices,
financial markets,
recommendation systems,
traffic forecasts.

The loop is:

PREDICTION

↓

BEHAVIOUR CHANGES

↓

SYSTEM CHANGES

↓

ORIGINAL PREDICTION MAY NO LONGER HOLD

These are reflexive systems.

A prediction can become an intervention

Suppose a bank predicts:

Applicant has high default risk.

It rejects the loan.

Now we never observe whether the applicant would actually have defaulted.

The prediction altered the path through reality.

So deployed prediction systems can change their own future datasets.

Prediction affects what becomes observable

The loop becomes:

MODEL

↓

PREDICTION

↓

DECISION

↓

WHO RECEIVES AN OPPORTUNITY

↓

OBSERVED OUTCOME

↓

NEW TRAINING DATA

This is why prediction systems cannot always be evaluated independently of the service in which they operate.

Predictions can be self-fulfilling

Suppose an influential model predicts:

Asset price will rise.

Traders buy.

Their buying contributes to:

price rise.

The prediction helped produce the outcome.

This is a self-fulfilling prediction.

Predictions can be self-defeating

Suppose a model predicts:

Electricity demand will be too high.

A system sends a price signal.

Consumers reduce demand.

The predicted shortage disappears.

Was the prediction wrong?

Not necessarily.

The forecast triggered an intervention that prevented the predicted outcome.

This distinction matters in control systems.

Successful prediction can erase its own evidence

Suppose a disease surveillance system predicts an outbreak.

Authorities intervene.

The outbreak never occurs.

Someone might later say:

The model predicted an outbreak and was wrong.

But perhaps:

prediction → intervention → outbreak prevented.

Evaluation becomes more complicated when predictions influence actions.

Control is prediction plus intervention

A controller repeatedly asks:

Where is the system likely to go?

Then:

What action moves it toward the desired state?

This is the connection between:

prediction

and:

control.

We will explore this more deeply when we reach state, dynamics and decision-making.

Prediction can be passive or active

Passive prediction

What will demand be tomorrow?

Action-conditional prediction

What will demand be if we set price to X?

The second is much closer to decision-making.

It asks about possible futures under different actions.

Counterfactual prediction

Suppose we have two possible actions:

and:

We want:

P(Future | Action A)

and:

P(Future | Action B).

Only one action will actually occur.

The other future remains counterfactual.

Decision-making requires comparing these unrealised possibilities.

Prediction and simulation

Prediction estimates what is likely to happen.

Simulation asks:

Given a model and assumptions, what happens if we run this scenario?

The two overlap.

But simulation can explore scenarios that are:

unlikely,
hypothetical.

For example:

What happens if demand increases by 50%?

That may not be the most likely forecast.

But it can still be valuable for planning.

Prediction and explanation

A model may predict very well without being easy to explain.

Suppose a neural network predicts equipment failure with:

99% accuracy.

But engineers cannot easily identify which physical mechanism it learned.

Is that acceptable?

It depends on the application.

For some decisions, predictive performance may be enough.

For others, explanation may be required for:

safety,
accountability,
trust.

Prediction and understanding are different

This is philosophically important.

Imagine an astronomical model that predicts planetary movement accurately.

We may also want a theory explaining why planets move that way.

Similarly, AI can discover predictive relationships without necessarily providing human-understandable explanations.

So:

prediction

and:

understanding

overlap but are not identical.

Better prediction can still create worse outcomes

Suppose an engagement algorithm becomes much better at predicting:

What will keep users watching?

Engagement rises dramatically.

But perhaps:

wellbeing falls,
misinformation spreads.

The prediction improved.

The service outcome worsened.

This illustrates again:

Predictive accuracy is not the same as good system design.

What should we predict?

Even choosing the target is a design decision.

Suppose a hospital wants to improve patient care.

Possible prediction targets include:

length of stay,
mortality,
likelihood of readmission,
expected treatment benefit.

Each target creates a different system.

Choosing the target determines what the machine pays attention to.

Labels define the prediction task

Suppose a hiring system predicts:

successful employee.

How is success defined?

promotion?
manager rating?
tenure?
sales?

The target label is not automatically given by nature.

Someone defined it.

That choice shapes what the model learns.

Prediction inherits the measurement system

Suppose:

employee success = manager rating.

The model predicts manager ratings.

If manager ratings contain:

bias,
inconsistency,

the model learns those patterns.

The prediction may be accurate relative to the label while failing to measure the deeper concept we actually care about.

Good prediction requires a good target

Before asking:

Which algorithm should we use?

ask:

What are we actually trying to predict, and why?

This is often the more important question.

A technically brilliant model solving the wrong prediction problem is not useful.

Prediction is time-bound

Suppose we predict:

customer default.

Over what period?

next month?
next year?
next ten years?

These are different targets.

Similarly:

machine failure in the next hour

is different from:

machine failure during the next year.

A prediction requires a defined horizon.

Prediction is space-bound

Suppose we predict:

electricity demand.

Where?

one house,
one neighbourhood,
one country?

Different spatial aggregation creates different prediction problems.

As we saw earlier:

time and space are part of the variable definition.

A useful prediction is contextual

Instead of:

predict electricity demand

we might define:

Predict half-hourly electricity demand for this network area 24 hours ahead.

Now we have:

target,
location,
time resolution,
forecast horizon.

The problem has become much clearer.

Forecast origin matters

Suppose we predict demand for:

18:00 tomorrow.

A prediction made at:

17:30 tomorrow

has much more current information than one made:

one week earlier.

So forecast quality depends on:

when the prediction is made.

We can call this the forecast origin.

Information available at prediction time matters

This is crucial for evaluating models.

Suppose a historical dataset includes tomorrow's temperature.

A model uses it to predict yesterday's demand.

Excellent accuracy.

But if that temperature forecast would not actually have been available at prediction time, the model is cheating.

This is data leakage.

Data leakage

Data leakage occurs when information unavailable at real prediction time enters the model during training or evaluation.

For example:

Predict whether a patient survives using a variable recorded after treatment ended.

The prediction may be accurate.

It is not useful.

The model has accidentally seen the future.

The past-present boundary must be respected

Recall our course timeline:

PAST → PRESENT → FUTURE.

When making a prediction at time t:

inputs must come from information available at or before t.

The target lies after that boundary.

This is especially important in time-series prediction.

Prediction requires a clean information boundary

Conceptually:

INFORMATION AVAILABLE NOW

│

│ ← prediction boundary

│

UNKNOWN OUTCOME

A valid predictive model must not allow information from beyond that boundary to leak backward.

Otherwise, evaluation becomes unrealistically optimistic.

Machine learning learns predictive functions

The core machine-learning task is often:

Find a function f that maps inputs X to target Y well enough to generalise.

So:

X → f → Ŷ

Training provides examples:

(x₁, y₁)

(x₂, y₂)

(x₃, y₃)

The algorithm adjusts the model so that:

f(x)

becomes a useful predictor of:

But what does "good" mean?

This immediately raises another question.

Suppose:

actual = 10

Model A predicts:

Model B predicts:

Model A seems better.

But how much better?

What if one kind of mistake matters much more than another?

Before a model can learn, we need a way of scoring prediction error.

That will lead to:

loss functions.

Different prediction tasks require different scoring

For numerical predictions, we might care about:

absolute error,
squared error.

For classification:

accuracy,
precision,
recall.

For probabilistic forecasts:

calibration,
proper scoring rules.

There is no single universal measure of prediction quality.

Error consequence matters

Suppose an autonomous vehicle predicts pedestrian presence.

False positive

It brakes unnecessarily.

False negative

It fails to brake for a pedestrian.

Both are classification errors.

Their consequences are radically different.

Prediction quality must eventually be judged relative to the service.

Predictions support different objectives

Suppose two models predict hospital demand.

Model A

Lower average error.

Model B

Better at predicting extreme demand peaks.

Which is better?

If the objective is:

average staffing efficiency

perhaps A.

If the objective is:

avoid capacity shortages

perhaps B.

There is no context-free best predictor.

Prediction becomes valuable when it changes action

Suppose we can predict something extremely accurately.

But no decision depends on it.

What is the value of the prediction?

Perhaps very little.

Prediction creates value when it allows us to make better decisions.

This leads to the value of information.

Perfect prediction can have zero value

Suppose you can perfectly predict:

What colour shirt someone will wear tomorrow.

If no decision depends on that information, it may have little practical value.

Now suppose you improve an electricity-demand forecast slightly.

That improvement allows:

less reserve,
fewer shortages.

A small predictive improvement may have enormous value.

Prediction value is decision-dependent.

Prediction can reduce resource requirements

Better demand forecasting may allow a service to hold less:

spare inventory,
reserve capacity,
staff.

Why?

Because uncertainty has decreased.

Prediction can therefore convert information into more efficient resource use.

But prediction cannot eliminate physical scarcity

Suppose we perfectly predict:

100 people will require treatment.

Only:

20 treatments exist.

Perfect prediction has not solved scarcity.

It has simply made the shortage certain.

Now the problem is:

allocation.

This is why prediction cannot be the endpoint of intelligent service design.

AI is often a prediction machine

One useful way to understand much of modern AI is:

AI dramatically reduces the cost of producing certain kinds of predictions.

Examples include:

image → object label

audio → transcript

text → next token

symptoms → disease probability

history → demand forecast

Many AI breakthroughs can be understood through this lens.

But modern AI is becoming more than prediction

A model can predict.

An agent can:

predict,
decide,
use tools,
act,
observe consequences.

A robot can do all of this in the physical world.

So the progression is:

PREDICTION

↓

DECISION

↓

AGENCY

↓

AUTONOMY

Prediction is the foundation.

It is not necessarily the endpoint.

Language models are prediction systems

At their core, large language models perform a remarkably simple prediction task:

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

Conceptually:

CONTEXT

↓

MODEL

↓

PROBABILITY DISTRIBUTION OVER NEXT TOKENS

↓

TOKEN SELECTED

↓

NEW CONTEXT

Then repeat.

Complex behaviour can emerge from simple prediction

This is one of the remarkable facts of modern AI.

Training a model to predict the next token across enormous amounts of data can produce capabilities involving:

writing,
translation,
coding,
summarisation,
question answering.

A seemingly modest prediction objective creates rich internal representations.

This will be explored much more later in the course.

Prediction forces representation

To predict what comes next in text, a model may benefit from representing things such as:

grammar,
concepts,
relationships,
context.

To predict future video frames, a system may need to represent:

objects,
motion,
physical structure.

Prediction can therefore encourage models to learn representations of the underlying world.

World models

Suppose an AI wants to predict:

What happens next if I take Action A?

It may develop an internal model of how the environment behaves.

This is sometimes called a world model.

Prediction becomes a mechanism for learning:

how states evolve.

This brings us towards intelligent agents.

Prediction of the next state

For a dynamic system, we can write conceptually:

current state + action + disturbance → next state.

Or:

x(t+1) = f(x(t), u(t), w(t))

where:

x(t) = current state,
u(t) = action,
w(t) = disturbance,
x(t+1) = next state.

Prediction asks:

Given where we are now and what happens next, where might we end up?

State prediction closes the course loop

Recall:

PAST

↓

DATA

↓

ESTIMATED PRESENT STATE

↓

MODEL

↓

POSSIBLE FUTURE STATES

Prediction is the bridge between:

present state

and:

possible futures.

Decision-making will later choose among actions that influence which future becomes real.

Prediction should preserve uncertainty

A weak framework is:

INPUT → ONE ANSWER

A stronger framework is:

INPUT

↓

MODEL

↓

PREDICTION + UNCERTAINTY

This can include:

expected value,
distribution,
confidence,
prediction interval.

We should not force uncertain worlds into falsely certain outputs.

Prediction is an information service

One useful way to think about prediction is as a service.

The service receives:

information available now.

It produces:

information about something not yet known.

The quality of that service depends on:

accuracy,
calibration,
timeliness,
cost,
reliability.

This perspective will become useful when we later design intelligent services.

A late prediction can be useless

Suppose a model predicts a machine will fail:

one second before failure.

Technically correct.

Operationally perhaps useless.

Predict:

one week beforehand

and maintenance may be scheduled.

Prediction value depends on timeliness.

The earliest prediction is not necessarily best

But predicting further ahead generally increases uncertainty.

So there is a trade-off:

earlier prediction → more time to act

but:

greater uncertainty.

This is another reason forecast horizon matters.

Prediction quality is multidimensional

A good prediction may need to be:

accurate,
calibrated,
timely,
robust,
useful.

A system that is excellent on one dimension may be poor on another.

For example:

extremely accurate but too late

may be less valuable than:

slightly less accurate but early enough to act.

The prediction problem comes before the algorithm

Before asking:

Should we use a neural network?

ask:

What is the target?
What information is available?
At what time?
At what location?
What horizon matters?
What uncertainty should be reported?
What decision will use the prediction?
What errors matter most?

Only then does model selection make sense.

Ask prediction questions

Whenever someone says:

We should use AI to predict X.

ask:

What exactly is X?
When is the prediction made?
What future horizon is being predicted?
At what spatial resolution?
What inputs are available at prediction time?
What historical examples exist?
Is the deployment world similar to the training world?
What uncertainty should be produced?
How will prediction quality be scored?
What happens when the model is wrong?
What decision uses the prediction?
Does the prediction itself change the system?

These questions turn a vague AI project into a well-defined prediction problem.

The central idea

Prediction is the process of using available information to estimate an outcome that has not yet been observed.

At its simplest:

INPUTS

↓

MODEL

↓

PREDICTION

But the richer version is:

PAST DATA

PRESENT CONDITIONS

↓

MODEL

↓

DISTRIBUTION OVER UNKNOWN OUTCOMES

↓

PREDICTION + UNCERTAINTY

↓

REAL OUTCOME

↓

ERROR

↓

LEARNING

Prediction connects:

what we know

to:

what we do not yet know.

It is one of the core functions of machine learning.

But prediction is not:

certainty,
causation,
optimisation,
decision-making.

It is information.

The next question is therefore:

What information goes into a prediction, and what exactly comes out?

That is the subject of the next lesson:

inputs and outputs.