Predictive Analytics
If you would do the same thing either way, the prediction is decoration.
What is predictive analytics?
Predictive analytics uses your historical data to estimate what is likely to happen next, and is worth building only where the estimate would change a decision that somebody is actually able to make.
Before any question about data or accuracy, there is a simpler test that most proposed predictions fail. If you knew the answer, would you do anything differently?
A prediction that customer demand will rise next month is interesting. If you cannot get more stock in that time and cannot add capacity, it is interesting and useless. The useful version predicts far enough ahead that you can still act.
The second test is data. Prediction needs enough history covering enough variation, and Kenyan businesses frequently have three years of records through a period that included nothing like normal conditions.
The test that removes most requests
Duara builds predictive work in Kenya only where a prediction would change a decision and the data can support it.
We apply the action test before the data test, and it removes most requests. A client who cannot act on the answer does not need a better model, and building one anyway produces an accurate number nobody uses.
Who this is for
- You want to anticipate rather than react
- A decision has a lead time you could exploit
- You have several years of consistent history
- Previous predictions were accurate and unused
Industries we serve
- Distribution and wholesale
- Manufacturing
- Retail chains
- Financial services
- Agriculture and agribusiness
- Logistics
Problems this solves
These describe the narrow conditions where a prediction is worth having rather than merely interesting.
- Predictions nobody can act on. The forecast is accurate and arrives too late to change anything, which makes it decoration.
- Insufficient history. Three years of records is frequently not enough, particularly where those years contained nothing resembling normal conditions.
- Accuracy treated as the goal. A model is judged on how close it was rather than on whether any decision changed because of it.
- Predictions presented as certainty. A range gets reported as a number, and the number gets treated as a fact by everybody downstream.
What we deliver
Applying the action test before building, checking whether history supports a prediction at all, expressing results as ranges rather than numbers, and measuring whether decisions actually changed.
An action test applied before anything is built
An honest assessment of whether history supports it
Predictions expressed as ranges, not single numbers
The lead time required to act, stated explicitly
Measurement of whether decisions changed
A recommendation against where appropriate
Features
- The action test comes first. If you would do the same thing either way, no model accuracy makes the prediction worth building.
- Lead time built into the design. A prediction has to arrive early enough for you to still do something, which frequently means predicting further out and less precisely.
- Ranges rather than numbers. A single figure gets treated as fact downstream, and reporting the range is what keeps the uncertainty visible.
- Honest about insufficient history. Where your data cannot support a prediction we say so, rather than producing a confident number from thin evidence.
- Measured on decisions, not accuracy. The test is whether anything was done differently, not how close the forecast turned out to be.
Benefits
- Money not spent on decoration. The action test removes most proposed predictions before anybody builds one.
- Predictions that arrive in time. Designing around lead time is what makes a forecast actionable rather than merely correct.
- Uncertainty stays visible. Ranges prevent an estimate being treated as a fact by everybody who receives it second hand.
- Better decisions where it applies. Where the conditions genuinely hold, anticipating rather than reacting is a real advantage.
Technologies we use
- Action testing. Establishing what decision would change and whether it can be made in time.
- History assessment. Judging whether your data covers enough variation to support a prediction.
- Forecasting. Producing estimates with ranges rather than single point figures.
- Lead time design. Predicting far enough ahead that action remains possible.
- Decision change measurement. Testing whether anybody actually did anything differently.
The most common failure is not inaccuracy. It is an accurate prediction arriving after the point at which anybody could have acted on it, which is a design failure rather than a modelling one.
How the process works, step by step
Predictive work usually takes six to twelve weeks where it proceeds, and the action test frequently ends it in the first week.
-
Step 1
Free action review
We ask what you would do differently if you knew, which removes most requests before any data discussion.
-
Step 2
Lead time established
How far ahead the prediction must arrive to be actionable, which shapes what can be built.
-
Step 3
History assessed
Whether your records cover enough variation, and we say plainly when they do not.
-
Step 4
Built with ranges
Estimates are produced as ranges so uncertainty survives being passed on.
-
Step 5
Decisions measured
We check whether anything actually changed, which is the real test.
Quality assurance and how we work
The commitments attached to telling a business what will happen next.
- The action test comes before the data question. A client who cannot act on the answer does not need a better model, and building one produces an accurate number nobody uses.
- We report ranges, not single figures. A number gets treated as fact by everybody downstream, and the range is what keeps the uncertainty attached to it.
- We say when your history cannot support it. Producing a confident forecast from thin evidence is worse than declining, because somebody will act on it.
- We measure whether decisions changed. Accuracy is not the test. Whether anything was done differently is, and that is frequently uncomfortable to report.
- We do not predict individual behaviour. Forecasting what a named person will do raises fairness questions we would rather not create for a client.
Frequently asked questions about predictive analytics
What is the action test?
If you knew the answer, would you do anything differently. Most proposed predictions fail it before any data question arises.
Why apply it before checking data?
Because a client who cannot act on the answer does not need a better model. Building one produces an accurate number nobody uses.
How much history do we need?
Enough to cover real variation, which is usually more than three years and depends entirely on how unusual those years were.
What if our history is too short?
We say so rather than producing a confident forecast from thin evidence. Somebody will act on that number and it will not deserve the confidence.
Why report ranges instead of numbers?
Because a single figure gets treated as fact by everybody who receives it second hand, and the uncertainty disappears at the first retelling.
Is accuracy not the goal?
No. The goal is a changed decision. An accurate prediction that arrives too late to act on is a design failure rather than a success.
What is the most common failure?
Not inaccuracy. It is an accurate forecast arriving after the point at which anybody could have done anything about it.
Can you predict customer demand?
Where history supports it and lead time allows action. Predicting a rise you cannot stock for is interesting and useless.
Can you predict which customers will leave?
Sometimes, and we are careful about acting on it. Predicting behaviour for named individuals raises fairness questions worth considering first.
Will you predict equipment failure?
Only with sensor data. Most Kenyan operations do not have it, and we would rather say so than sell a model that cannot work.
How long does predictive work take?
Six to twelve weeks where it proceeds, and the action test frequently ends the conversation in the first week.
What counts as a successful forecast?
One that changed a decision, not one that turned out close. That is frequently uncomfortable to report and it is the honest measure.
What about seasonality?
Learned from your own history where the history covers full cycles. Where it does not, we say the seasonal component is unsupported.
Can it handle sudden changes?
Poorly, and no forecasting method handles them well. A model trained on stable conditions will be confidently wrong through a disruption.
Does it replace judgement?
No. It informs it, and where a prediction disagrees with somebody experienced, the disagreement is usually worth investigating rather than resolving in favour of the model.
Can we start small?
Yes, and one prediction against one decision is the right size. Broad forecasting programmes are where this work usually goes wrong.
What data do you need?
Your own operational history. Buying external data rarely helps as much as suppliers of external data suggest.
Will the model need retraining?
Yes, and conditions change, so a forecast that was accurate last year is not automatically accurate now.
Who sees our historical data?
Only the people working on your project, and it stays in your systems.
What does predictive work cost?
Data condition and decision complexity determine it, quoted after the action review, which may recommend against proceeding.
Glossary of terms
Forecasting terms in plain language.
- Action test
- Whether you would do anything differently if you knew the answer, applied before any model is built.
- Lead time
- How far ahead a prediction must arrive for action to still be possible.
- Range
- The band a prediction falls within, which should be reported instead of a single figure.
- Historical variation
- Whether your records cover enough different conditions to support a forecast.
- Point estimate
- A single predicted number, which downstream readers treat as certainty.
- Decision change
- Whether anybody acted differently because of a prediction, which is the only real measure.
Related services in AI Solutions
Talk to us about predictive analytics
Take the prediction you most want and ask what you would do the day you received it. If the answer is nothing, that settles it.
Start your project