The Journey of Artur Linx: A Financial Analyst Navigating the World of Machine Learning

In the world of corporate finance, much is said about revenue projections, demand forecasts, and budget planning. But behind the complex spreadsheets and endless meetings lies a question that is rarely asked in depth:

“Do our forecasts truly reflect the reality of the business?”

It was this very question that led Artur Linx, a senior financial analyst, to cross the border between traditional finance and the world of machine learning.

Artur was not a data scientist. He was a seasoned professional in income statements, EBITDA margins, seasonality, and volume variations. But over the past few years, he realized something important: technology didn’t have to be just an operational ally — it could become an extension of financial analytical thinking.

When studying machine learning, his goal wasn’t to automate his work, but to enhance his ability to interpret business data. He understood that sophisticated predictive models required more than artificial intelligence — they required business context, something he deeply understood.

His challenge was concrete: the company’s demand forecasts were inconsistent. This had a direct impact on working capital, leading to the maintenance of excess inventory to compensate for unpredictability and maintain service levels — which in turn strained logistical planning and the company’s financial health.

One evening, coffee in hand, Artur asked himself a new question:

“Do I have the right data? Is it structured?”

Reviewing the current process, he realized the forecasts were based solely on last year’s prices and volumes, adjusted through informal conversations with sales and marketing. It was time for a change.

He began by expanding the scope: he collected data from the past five years on price and volume by region and sales channel — structured and accessible information. But he knew this alone wasn’t enough. To truly understand demand behavior, he needed to enrich the dataset.

He sought data from marketing campaigns and promotions. He discovered that this data was scattered across unorganized spreadsheets, emails, and often in free-text formats. After a considerable effort in organizing and cleaning the data, he managed to incorporate it. He also added commercial dates and relevant holidays. Finally, he had a robust database ready for model testing.

Like any good financial analyst, Artur started with what he knew: Excel. He applied a linear regression — simple but effective to get started. However, the results showed significant errors between predicted and actual values. It was time to level up.

He recalled that his company had recently implemented a new planning system that supported parameterized machine learning models — no coding required. Although he knew how to code in Python, Artur made a strategic decision:

“I’ll use the company’s system to ensure scalability and replicability.”

He began by testing three algorithms — simply by selecting them in the software:

• Decision Tree: average error of 9.07%

• Random Forest: 8.51%

• XGBoost: 6.8%

Excited by the results, he considered further improvements. He realized that marketing actions didn’t necessarily impact demand in the same month they were executed. He adjusted the timing of campaigns, testing lag effects. After reprocessing the models, the errors dropped

• Decision Tree: 8.05%

• Random Forest: 7.51%

• XGBoost: 5.8%

Still, Artur wasn’t satisfied. What else could influence demand?

He had an insight: the weather. Temperature and rainfall could affect demand for certain products. He accessed meteorological databases, collected five years of historical weather data, and integrated it into the model.

With this new data, the results improved even more — now the forecasts were finally aligned with business reality. Artur now had a solid model to collaborate with the Marketing and Sales teams. He immediately sent an email to schedule a meeting.

Key Learnings from Artur Linx’s Journey with Machine Learning

1.Data is the starting point

No machine learning technique works without a solid data foundation. Having structured, clean, and reliable data is the essential first step for any predictive analysis.

2.Business knowledge is indispensable

Analytical tools lose their power without an understanding of the context in which they operate. Machine learning without business acumen becomes mere statistical experimentation — with little real impact.

3.Modeling is an iterative process

Building a good model requires patience and method. Testing, adjusting, enriching data, and reevaluating are part of the natural cycle. And more: models need to evolve over time to keep up with market changes.

4.Accessible technology expands adoption

With the advancement of planning and analytics systems, knowing how to code is no longer a requirement for using machine learning. Modern software allows users to parameterize complex models through intuitive interfaces, democratizing access to analytical intelligence.