Forecasting: Methods, Accuracy & Applications

TAKEAWAYS:

• Forecasting predicts future demand, supply, or trends using quantitative (data-driven) or qualitative (expert-based) methods.
• Time-series models (naïve, moving average, exponential smoothing) rely on historical data, while causal models (regression) link demand to external factors.
• Forecast accuracy is measured by MAD, MSE, MAPE, and tracking signals—critical for inventory and production planning.
• Qualitative methods (Delphi, jury of executive opinion) dominate when data is scarce (e.g., new product launches).
• Bias and error in forecasts arise from random variation, systematic bias, or structural breaks—mitigated via robust validation.
• Real-world applications include Daraz’s demand forecasting for e-commerce, Nabil Bank’s loan default prediction, and NTC’s subscriber growth modeling.

---

1. What Is Forecasting?

Forecasting is the art and science of predicting future events (e.g., demand, sales, costs, or resource needs) to support decision-making in operations, finance, and marketing. It bridges the gap between current data and future uncertainty, enabling businesses to:

• Optimize inventory levels (avoid stockouts or excess).
• Plan production schedules (e.g., Daraz’s warehouse orders).
• Allocate resources efficiently (e.g., NTC’s network capacity).
• Set pricing and promotions (e.g., Pathao’s surge pricing).

Key Terms:

---

2. Why Is Forecasting Critical?

How forecasting ties supply chain stages (Image: Miguel Garcia Gonzalez, CC BY-SA 4.0, via Wikimedia Commons)

Real-World Impact:

• Daraz (Nepal): Uses machine learning forecasting to predict demand for electronics (e.g., laptops during exam seasons) and adjusts warehouse stock accordingly.
• Nabil Bank: Forecasts loan defaults using historical repayment data and economic indicators (e.g., inflation rates).
• NTC: Models subscriber growth to expand 4G/5G infrastructure in underserved regions.

Case Study: Kathmandu Traffic Congestion The Kathmandu Metropolitan City uses short-term traffic flow forecasting to:

1. Predict peak hours (e.g., 7–9 AM on weekdays).
2. Adjust traffic light timings dynamically (via smart sensors).
3. Reduce average travel time by 15% (source: KMC Smart City Project, 2022). Worked Example: If historical data shows 30% more vehicles on Thursdays, the city schedules extra buses and police patrols to mitigate gridlock.

---

3. Types of Forecasting Methods

Forecasting methods are classified into two broad categories:

A. Qualitative Methods (Judgment-Based)

Used when data is scarce, new products/services are launched, or expert opinion is critical.

Advantages:

• Useful for long-term or one-time events.
• Incorporates expertise and local knowledge.

Disadvantages:

• Subjective (biased by opinions).
• No data-backed accuracy metrics.

---

B. Quantitative Methods (Data-Driven)

Relies on historical data, statistical models, or machine learning. Subdivided into:

1. Time-Series Forecasting

Assumes future values depend on past values (e.g., monthly sales). Models include:

Worked Example: Exponential Smoothing for Daraz Suppose Daraz’s monthly demand for smartphones (in units) is:

Interpretation:

• The forecast adjusts closer to actual demand as new data arrives.
• α=0.3 means past forecasts matter more than recent spikes (avoids overreacting to noise).

2. Causal Models

Assumes demand depends on external factors (e.g., price, advertising, holidays).

Worked Example: Regression for Nabil Bank Nabil Bank wants to predict loan defaults based on:

• $X_1$: Customer income
• $X_2$: Loan amount
• $Y$: Default (1 = yes, 0 = no)

Model Output: $Y=0.5-0.002X_1+0.005X_2$

• If a customer earns ₹500,000 and takes a ₹2,000,000 loan: $Y=0.5-0.002(500)+0.005(2000)=0.5-1+10=9.5$ → Logistic regression would convert this to a probability (e.g., 80% default risk).

---

4. Measuring Forecast Accuracy

No forecast is perfect. Error metrics quantify performance:

Worked Example: Evaluating a Forecast for Khalti Khalti’s monthly transaction volume (in millions):

Calculations:

• MAD = (10 + 5 + 20)/3 = 11.67 million
• MAPE = (8.3 + 3.8 + 13.3)/3 = 8.5%
• Tracking Signal = (10 + 5 + 20)/(10 + 5 + 20) = 1 (no bias, but high error).

Insight:

• The forecast underestimates March demand (likely due to Dashain festival spending).
• Khalti should adjust α in exponential smoothing or add a holiday factor.

---

5. Challenges and Limitations

---

6. Advanced Techniques

For high-stakes forecasting (e.g., supply chains, finance), businesses use:

1. Machine Learning (ML) Models

   • Random Forests: Handles non-linear relationships (e.g., NTC’s network traffic).
   • ARIMA: Captures autocorrelation in time-series (e.g., stock prices).
   • Neural Networks: Predicts complex patterns (e.g., Google’s demand forecasting).

2. Collaborative Forecasting

   • Cross-functional teams (sales, operations, marketing) align forecasts.
   • Example: Unilever Nepal uses S&OP (Sales & Operations Planning) to sync production with retail demand.

3. Simulation and Scenario Planning

   • Monte Carlo Analysis: Models probabilistic outcomes (e.g., "What if demand drops 20%?").
   • Example: Nabil Bank simulates interest rate hikes to stress-test loan portfolios.

---

In the Real World

1. Daraz (Nepal)

   • Idea Used: Time-series forecasting + machine learning
   • How: Daraz’s demand planning team uses Python-based ARIMA models to predict:
     • Product-specific demand (e.g., laptops spike before exams).
     • Warehouse stock levels (avoids overstocking perishables like milk).
   • Impact: Reduced stockout costs by 25% and warehouse holding costs by 18% (2022 report).

2. Nabil Bank (Nepal)

   • Idea Used: Regression analysis + qualitative judgment
   • How: Predicts loan defaults using:
     • Quantitative: Customer income, loan amount, credit score.
     • Qualitative: Economic outlook (e.g., "Will inflation rise in 2024?").
   • Impact: Improved default prediction accuracy by 30%, reducing bad loans.

3. NTC (Nepal Telecom)

   • Idea Used: Exponential smoothing + causal models
   • How: Forecasts:
     • Daily call volumes (naïve/moving average for short-term).
     • 4G/5G subscriber growth (regression: income vs. adoption).
   • Impact: Optimized network capacity expansion in Kathmandu and Pokhara.

4. Pathao (Nepal)

   • Idea Used: Real-time demand forecasting
   • How: Uses surge pricing algorithms (like Uber) based on:
     • Time of day (peak hours = higher prices).
     • Weather data (rain = more demand for rides).
   • Impact: 20% higher driver earnings during peak times.

---

Exam Tip

How This Unit Is Tested (PU Pattern)

1. Definitions & Concepts (20%)

   • Expect short-answer questions on:
     • Difference between qualitative vs. quantitative forecasting.
     • Explain exponential smoothing with α.
     • Define MAD, MAPE, tracking signal.

2. Problem-Solving (40%)

   • Worked examples will ask you to:
     • Calculate a 3-month moving average.
     • Compute MAPE for given data.
     • Choose the best forecasting method for a scenario (e.g., "New product launch with no historical data → Delphi method").

3. Case Studies (30%)

   • Apply forecasting to real businesses:
     • "How would Daraz forecast demand for Diwali sales?" → Exponential smoothing + promotion factors.
     • "NTC wants to predict 5G adoption. Which method?" → Regression (income vs. adoption) + qualitative (government policies).

4. Critical Analysis (10%)

   • Discuss limitations of a method (e.g., "Why might exponential smoothing fail?" → Structural breaks, high seasonality).

Common Mistakes to Avoid:

• Ignoring seasonality/trends in time-series data.
• Using the wrong α in exponential smoothing (too high = overreacts to noise; too low = slow to adapt).
• Not interpreting errors (e.g., high MAPE doesn’t always mean a bad forecast if actuals are volatile).

Quick Revision Table:

---

Final Note: Forecasting is not about perfection—it’s about reducing uncertainty to make better decisions. Whether you’re optimizing Daraz’s inventory, Nabil Bank’s loans, or NTC’s network, the right method + continuous error tracking will give you an edge.

Practice: Solve 10 past PU questions on forecasting—focus on calculations (MAD, MAPE) and method selection. Good luck!