“Finance has years of invoices, sales has the CRM, marketing has every campaign we have ever run. And when orders jump in March, nobody can say why.”
Data in every department, read together in none: the data exists, spread across departments and formats. Nobody reads it together.
What nuclai does in analysis and forecasting
At nuclai, analysis and forecasting means building predictive models from a company’s own history, after preparing that history so it can be trusted.
We first agree with you on what to predict: next quarter’s sales, the effect of a campaign, which customers are likely to leave. We take stock of the data you actually hold, then prepare and cross-reference it across finance, sales and marketing. Only then do we build the model and put its output in the hands of the teams who decide. The quality of a forecast depends first on that preparation.
Data preparation: the step that decides the forecast
A model learns only from what it is given.
- Selecting the useful data. We keep the sources that bear on the question and set the rest aside, with the reason written down.
- Organising it. Customer IDs that differ between the CRM and accounting, product codes that changed, periods that do not line up: we reconcile them so each record means the same thing everywhere.
- Cross-referencing departments. Invoices say what was bought, the CRM says by whom, the campaign calendar says what was running at the time. Read together, they describe customer behaviour in a way none of them does alone.
- Weighting each input. Recent years may count more than older ones; a one-off contract may need to count less. Each weighting is a choice, documented and reviewed with you.
- Separating events from the trend. A campaign, a promotion, a stock shortage, an exceptional order: each is logged as an explained event, so the model does not mistake it for the underlying trend.
This is where your questions become answerable. Did the spring campaign cause the March spike, or did it coincide with a seasonal peak that comes back every year? Do you know who your ideal customer is, how they buy, and which of your campaigns actually brought them to a purchase? Cross-referenced data can help answer with evidence. Data kept apart cannot.
There are limits: a model extends the patterns in your history, and gives an estimate with a margin of error.
What you receive
At the end of the engagement, you hold five things:
- The forecasting brief: the questions retained, how each is defined and measured, and the horizon of each forecast.
- The data map and preparation rules: sources used and set aside, links between departments, weightings, and the log of explained events.
- The preparation scripts and the models, installed in your environment, which your team can re-run as new data comes in.
- A forecast view for the teams, format of the output can and should be decided with you, e. g. view in the existing BI tool, a dynamic spreadsheet or dedicated dashboard.
- The handover pack: model documentation (inputs, assumptions, limits, margin of error), user guide and any prompts used, so the models run with the tools or LLMs of your choice.
All five are stored on your side.
Four steps: diagnostic, pilot, rollout, handover
- Diagnostic. The financial process diagnostic runs in three phases: on site, analysis, debrief. For forecasting, it establishes what you want to predict, which data can support it, and whether a model is worth building. It may conclude that it is not. It is covered by its own proposal, and its flat fee is deducted in full from the engagement if you go ahead.
- Pilot. We build a model for one forecasting question, chosen with you. It is first tested on past periods whose outcome you already know.
- Rollout. Once your team has signed off the pilot, the model goes into use, and further questions are added in the order the diagnostic set.
- Handover. We hand over documentation, scripts and access. From then on, the models are yours.
Duration depends on the scope set in the proposal: the number of questions, the number of sources, and the state of the data.
Who it is for
- Companies: SMEs and mid-sized companies with several years of sales and invoicing history.
- Roles: CFO, head of finance, financial controller, marketing director, head of sales.
- Prerequisites: accessible exports from your accounting system or ERP and from your CRM; a record of past campaigns and promotions, even an incomplete one; someone in each department involved who knows how its data was recorded; a completed diagnostic.
What changes, dimension by dimension
| Dimension | Before | After |
|---|---|---|
| Decisions | Taken on intuition and last year’s figures | Taken on a forecast with a stated margin of error |
| Campaigns | Results read campaign by campaign | Each campaign’s effect separated from season and trend |
| Ideal customer | Described from CRM segments and experience | Described from purchase behaviour across CRM and invoicing |
| Planning | Reacting once a shift shows in the figures | Anticipating, within the horizon the model supports |
| Ownership | Analysis held by one person or an external tool | Data rules and models documented and stored on your side |
Frequently asked questions about forecasting models
It depends on the question you want answered. A seasonal pattern needs several full cycles; a campaign effect needs enough campaigns to compare. The diagnostic sets this out for your case before any model is built, and it can conclude that you are not there yet.
A BI dashboard shows what has happened; a forecasting model estimates what is likely to happen next and which factors drive it. The forecasts can feed the dashboard your teams already use.
Customer behaviour can often be analysed without identifying individual customers. Which personal data is used, and in what form, is decided during scoping, including whether aggregated or pseudonymised data is enough. The data stays in your environment, under your control.
If the data is insufficient, no model is built, and you know it before committing to one. The diagnostic then tells you what to start recording, and how, so the question can be answered later.
A model can separate a campaign’s effect from seasonality and the underlying trend, provided campaigns are recorded alongside sales and invoicing. When campaigns overlap or are too few to compare, we say so rather than guess.
A forecasting model cannot predict events that have no precedent in your data. It extends the patterns of your history, so a new competitor, a regulatory change or a supply shock falls outside it. Every forecast comes with a documented margin of error, so your teams know how far to rely on it.
Request a proposal for a forecasting model
Every forecasting engagement starts with the financial process diagnostic: it establishes what your data can predict, and whether a model is worth building. The first conversation, remote and free of charge, sets its scope.
