Problem Definition

The first step in data science is to have a clear and concise problem definition to scope the task at hand.

The following table is a suggested framework to be used for the same:

Section	Example Entry
Context	Retail chain struggles with product stockouts causing lost sales.
Business Objective	Reduce lost sales by improving inventory planning.
Data Science Objective	Forecast daily product demand per store for next 14 days.
Key Questions	What is expected sales volume per product-store-day?
Success Metric	MAPE ≤ 15% for top 100 products.
Constraints	Must run nightly and integrate with existing PostgreSQL system.
Data Needed	Sales history, calendar events, weather, promotions.
Deliverables	Forecast table + visualization dashboard in BI tool.

This can be augmented by using a questions-to-datamap mapping in the following way:

Key Question	Data Needed	Data Source	Notes (Quality, Processing, Granularity)
What will demand be per store next week?	historical sales by store/date	sales database	aggregate to daily level
Does weather affect demand?	temperature, rainfall	weather API	join on geo + date
Do promotions change buying patterns?	promo flags, discount %, marketing schedule	marketing calendar	need categorical encoding

When preparing the problem definition, it is necessary to identify the analysis performance metrics to start with. The following provides an example of the same:

Analysis Type	Example Metrics	Interpretation Focus
Forecasting	MAE, RMSE, MAPE	Forecast error relative to scale
Regression	R², RMSE, MAE	Variance explained & error magnitude
Classification	F1 score, ROC AUC, Accuracy	Balance error vs. precision/recall trade-offs
Clustering	Silhouette Score, Davies-Bouldin	How well data points cluster
EDA Insights	Summary stats, distributions	Whether patterns are meaningful and reproducible

The conclusion stability also needs to be tested as an extension. To choose the method for it, we have the following:

Stress Test	How
Train/Test or Cross-validation consistency	k-fold or rolling window validation
Sensitivity analysis	Vary assumptions or thresholds
Stability across subgroups	Performance by region/customer/segment/time

A proper validation can be done using the following checkbox:

[ ] The analysis answers the original objective.
[ ] Data used is appropriate and clean, with no leakage.
[ ] Performance was assessed using correct evaluation metrics.
[ ] A baseline comparison was included.
[ ] Model/analysis is robust across tests and subsets.
[ ] Results are interpretable to stakeholders.
[ ] Actionable recommendations are clearly stated.
[ ] Ethical and contextual impacts were considered.

The thereotical aspects, such as the analysis performance metrics, can be found in the statistics section.