The Naked Number

Volume II · Forecasting & financial analytics

The Naked Number

A single figure, presented as certainty and planned against as fact.

The problem

You need to decide something now that depends on what happens later — how much to hold, whether cash covers the next two quarters, whether a price change pays for itself. The spreadsheet that answers it produces one number, and nobody can say how wrong that number could be.

Who it is for

Finance and operations teams making recurring decisions under uncertainty: demand and cash planning, pricing, capacity, scenario work for a board.

What you end up with

A model you can interrogate, compared against the simple baseline it has to beat, with the uncertainty stated rather than implied.

• A forecast with intervals, and a validation showing how it performed on periods it never saw.
• A comparison against a naive baseline, because a model that cannot beat last month’s number is not worth maintaining.
• The assumptions written down as parameters you can change, not constants buried in code.

What you end up with

• A recommendation on whether to build further, keep it simple, or stop.

Evidence record

The portfolio optimizer on this page is the method made inspectable: assumptions visible, views adjustable, and outcomes shown as a distribution rather than a projection. It runs on invented capital-market assumptions, which is stated on its face — it demonstrates how I build, not what a market will do.

Process · 01

We identify the decision the forecast serves and how much accuracy it actually needs. Often less than expected.

Discovery

The decision, its cadence, and the accuracy that would change it — written down.

Process · 02

I establish what a naive forecast achieves on your history, then test whether anything more elaborate beats it.

Baseline & assessment

A baseline comparison, a validation approach, and a recommended next step.

Process · 03

The chosen model is built where your team can run it, with the intervals surfaced rather than hidden.

Delivery

The model, its interface, and documentation of every assumption in it.

Process · 04

Re-validation is part of the handover: how to tell when the model has stopped working.

Handover

Runbook, re-validation procedure, and the monitoring to support it.

Data provenance

• Historical data at the grain the decision is made at, with its known gaps.
• The decision itself and who makes it.
• Any structural rules the outcome passes through — tax bands, tiers, covenants, thresholds.
• Known one-off events in the history, so they are not learned as patterns.

What is in scope

• Time-series forecasting, scenario modelling, sensitivity and stress analysis.
• Baseline comparison and out-of-sample validation.
• Financial models with the assumptions exposed.

Limitations

• Predictions of market prices, and any suggestion that a forecast removes uncertainty.
• Investment advice, portfolio management, or trading of any kind.
• Guaranteed accuracy. What is achievable is established by validation, not promised in advance.

Have a problem shaped like this?

Forecasting assessment

A baseline comparison, a validation approach, and a recommended next step.

How accurate will the forecast be?

Not knowable until it has been validated on your history. The assessment exists to answer this honestly, and sometimes the answer is that a simple baseline is already good enough and you should not pay for more.

We have messy, incomplete history. Is that disqualifying?

No, but it changes what is achievable and you should hear that early. Part of the assessment is establishing what the data will and will not support.

Can you forecast our share price or a market?

No. I do not take that work, and anyone who offers it with confidence is telling you something about themselves.

Why do you insist on showing a range?

Because the range is the decision-relevant part. A single number invites a plan with no contingency in it.

Approach

• Python
• pandas
• statsmodels
• scikit-learn
• PyTorch
• SQL

Making a decision that depends on a forecast?

Describe the decision and the data you hold, and I will tell you what is realistically achievable.