Background
This project was completed during my part-time role at Siemens Turkey's Regional Process Office, where I worked with operational sales data to examine how historical demand patterns could be translated into forward-looking information for business planning.
Using approximately five years of sales data across around 500 electrical products, I developed a time-series forecasting workflow designed to identify recurring demand patterns, generate forward-looking sales estimates, and communicate the results in a form that could support operational stakeholders.
The project connected statistical forecasting with a practical business question: how can historical demand data be used to support planning when future sales remain uncertain?
Research Goals
- Identify trend, seasonality, and recurring demand patterns in historical product-sales data.
- Develop a statistical forecasting model capable of representing seasonal sales behaviour.
- Evaluate forecast behaviour and residual structure rather than relying only on fitted values.
- Translate model outputs into interpretable forecasts and visualisations for business stakeholders.
- Integrate forecasting results into the broader reporting workflow used for operational planning.
Dataset
- History: Approximately 5 years of sales data
- Scope: Around 500 electrical products
- Frequency: Monthly sales observations
- Source: Internal operational sales data
- Key challenge: Substantial variation in demand patterns across products and over time
The dataset required cleaning, aggregation, and exploratory analysis before forecasting. Particular attention was given to recurring seasonal patterns and abrupt changes in product-level sales behaviour.
Methodology
- Time-Series Decomposition: Examined historical sales data to separate overall trend and recurring seasonal behaviour.
- SARIMA: Used Seasonal ARIMA models to represent both temporal dependence and recurring sales seasonality.
- Model Selection: Compared candidate parameter configurations to identify suitable autoregressive, differencing, moving-average, and seasonal components.
- Residual Analysis: Examined residual behaviour to assess whether major temporal patterns remained unexplained by the fitted model.
- Forecast Visualisation: Produced forward-looking forecasts together with historical sales patterns for interpretation and reporting.
- Tools: Python, pandas, statsmodels, SQL, Power BI
Results
The analysis showed that electrical-product sales contained clear seasonal structure alongside substantial product-level variation and occasional abrupt shifts in demand.
SARIMA provided a structured way to represent recurring seasonal behaviour and temporal dependence while generating forward-looking demand estimates.
The resulting forecasts and visualisations were presented to departmental management, allowing the statistical results to be considered alongside operational knowledge rather than treated as standalone model outputs.
The forecasting work was also incorporated into a broader live Power BI reporting workflow tracking product sales, customer activity, and demand.
Contributions
- Analysed approximately five years of operational sales data across around 500 electrical products.
- Developed a SARIMA-based forecasting workflow for seasonal product demand.
- Combined statistical modelling with exploratory analysis and residual diagnostics to examine model behaviour.
- Translated forecasting outputs into management-facing visualisations and reporting rather than limiting the project to offline model evaluation.
- Contributed to a live Power BI dashboard workflow for continued monitoring of sales and customer activity.
- Documented and transferred the reporting workflow so it could continue to be used after the project.
Impact
This project was my first experience applying forecasting directly within an operational decision environment.
The central challenge was not simply producing a numerical prediction. Product demand varied across time and categories, while planning decisions still had to be made before future demand was known. Forecasts therefore served as decision-support inputs for demand planning and reporting, rather than deterministic answers.
The experience later influenced my interest in AI-enabled decision-making under uncertainty by showing that the practical value of a predictive model depends not only on its statistical performance, but also on how its outputs are interpreted, communicated, and incorporated into real decision processes.