Business Analytics & Predictive Modelling: Techniques, Tools & Applications

Core Concepts: What is Business Analytics?

Business Analytics (BA) transforms raw data into actionable insights through statistical analysis, modeling, and optimization. It bridges the gap between data and decision-making by answering three key questions:

1. Descriptive Analytics: "What happened?" (e.g., sales trends, customer behavior)
2. Predictive Analytics: "What will happen?" (e.g., demand forecasting, churn prediction)
3. Prescriptive Analytics: "What should we do?" (e.g., dynamic pricing, route optimization)

mindmap
  root((Business Analytics))
    Descriptive
      "Summarizes past data (e.g., reports, dashboards)"
      Tools: SQL, Excel, Tableau
    Predictive
      "Uses ML to forecast outcomes (e.g., 'Will this customer buy again?')"
      Models: Regression, Decision Trees, Neural Networks
    Prescriptive
      "Recommends optimal actions (e.g., 'Set price at $X to maximize profit')"
      Techniques: Optimization, Simulation, AI

How It Works: The Analytics Pipeline

1. Data Collection: Gather structured/unstructured data (e.g., transaction logs, social media, sensors).
2. Data Cleaning: Handle missing values, outliers, and noise (e.g., using Python’s pandas or SQL’s CASE WHEN).
3. Model Training: Apply algorithms (e.g., linear regression for sales forecasting).
4. Validation: Test accuracy (e.g., confusion matrix for classification).
5. Deployment: Integrate into business processes (e.g., automated alerts for fraud).

Predictive Modelling: Algorithms and Applications

Predictive models use historical data to forecast future events. Key techniques include:

1. Regression Analysis

Definition: Predicts continuous outcomes (e.g., sales, temperature) using linear/non-linear relationships. Example: Nepal Electricity Authority (NEA) uses regression to forecast electricity demand based on weather and past usage.

Worked Example: Daraz’s Inventory Forecasting

• Problem: Overstocking leads to waste; understocking loses sales.
• Solution: Daraz trains a multiple linear regression model on:
  • Historical sales data (dependent variable: units_sold)
  • Independent variables: season, promotions, competitor_prices, holidays.
• Outcome: Reduces excess inventory by 15% by adjusting orders dynamically.

2. Classification Models

Definition: Predicts categorical outcomes (e.g., "Will a loan default?" or "Is this email spam?"). Algorithms:

• Decision Trees: Simple, interpretable (e.g., "If income > $50k AND credit_score > 700 → Approve Loan").
• Random Forest: Ensemble of trees for higher accuracy.
• Logistic Regression: Probabilistic classification (e.g., "85% chance of churn").

Real-World Use: Nabil Bank’s Loan Approval System

• Problem: Manual loan approvals were slow and biased.
• Solution: Nabil Bank built a Random Forest classifier trained on:
  • Applicant data: income, employment_history, credit_score, loan_amount.
  • Outcome: default (1) or no_default (0).
• Result: Approval time reduced from 3 days → 2 hours; default rate dropped by 12%.

3. Clustering (Unsupervised Learning)

Definition: Groups similar data points without predefined labels (e.g., customer segmentation). Algorithms:

• K-Means: Partitions data into K clusters (e.g., "Find 4 customer segments").
• Hierarchical Clustering: Builds a tree of clusters (dendrogram).

Example: Pathao’s Driver Optimization

• Problem: Drivers in Kathmandu had uneven earnings due to route imbalances.
• Solution: Pathao used K-Means clustering to group:
  • Cluster 1: High-demand zones (e.g., Thapathali, Lakshmi Marg).
  • Cluster 2: Low-demand zones (e.g., rural areas).
• Action: Dynamically adjusted surge pricing and driver incentives.
• Impact: Driver earnings increased by 20% in high-demand clusters.

4. Time-Series Forecasting

Definition: Predicts future values based on past trends (e.g., stock prices, weather). Models:

• ARIMA: AutoRegressive Integrated Moving Average (e.g., "Forecast NEPSE index").
• Prophet (Facebook): Handles seasonality (e.g., "Predict eSewa transactions during Dashain").

Case Study: eSewa’s Holiday Traffic Prediction

• Problem: Server crashes during Dashain due to unexpected user spikes.
• Solution: eSewa used ARIMA to forecast:
  • Dependent Variable: daily_transactions.
  • Independent Variables: day_of_week, holiday_flag, promotions.
• Outcome: Scaled servers dynamically, reducing downtime by 90%.

Business Analytics in Nepal: Case Studies

Exam Tip: How to Score Full Marks

1. Define Clearly: Always start with definitions (e.g., "Predictive modelling uses historical data to forecast future probabilities using algorithms like regression or decision trees.").
2. Use Real Examples: Examiners love Nepalese cases. Link models to companies (e.g., "Like Nabil Bank’s loan approval system, a retail firm could use Random Forest to...").
3. Show Maths Where Needed:
   • For regression, write the equation and interpret coefficients.
   • For classification, draw a confusion matrix and calculate accuracy.
4. Compare Models: Use tables to contrast techniques (e.g., "Decision Trees vs. Random Forest").
5. Visualize: Sketch a decision tree, time-series graph, or cluster plot in your answer. Even a rough diagram earns partial credit.
6. Link to Business Value: End each answer with an impact statement (e.g., "This reduces costs by X% and improves Y by Z%").

Common Pitfalls to Avoid:

• Confusing descriptive (what happened) with predictive (what will happen) analytics.
• Forgetting to validate models (always mention "train-test split" or "cross-validation").
• Overcomplicating answers—stick to 1–2 models per question unless asked for a comparison.