Forecasting Techniques, Models & Applications in OM

TAKEAWAYS

• Forecasting is the backbone of OM decisions: it predicts demand, supply, and trends to optimize inventory, capacity, and scheduling.
• Qualitative methods (judgmental) rely on expert opinions (e.g., market research), while quantitative methods use historical data (e.g., moving averages, regression).
• Time-series models decompose data into trend, seasonality, and random fluctuations—critical for seasonal businesses like tourism or agriculture.
• Causal models (e.g., regression) link demand to external factors (e.g., income, weather), used by banks for loan defaults or NTC for electricity demand.
• Forecast accuracy is measured by MAD, MSE, MAPE—lower errors mean better inventory/supply chain efficiency.
• Bias and random errors must be minimized; techniques like exponential smoothing adapt to recent trends dynamically.

1. What is Forecasting? Why Does It Matter?

Forecasting is the art and science of predicting future events based on past data, expert judgment, or market trends. In Operations Management (OM), it drives:

• Inventory planning (e.g., Daraz stocking products before Dashain).
• Capacity decisions (e.g., NTC adding power plants during summer).
• Supply chain coordination (e.g., banks forecasting loan demand).
• Risk management (e.g., Pathao predicting driver shortages).

flowchart TD
    A["Demand Forecasting"] --> B["Supplier Lead Time"]
    A --> C["Production Planning"]
    A --> D["Inventory Levels"]
    B --> E["Procurement"]
    C --> F["Workforce Scheduling"]
    D --> G["Warehouse Management"]
    E --> H["Delivery Logistics"]
    F --> I["Customer Orders"]
    G --> I
    H --> I

Real-World Example: NTC’s Load Forecasting Nepal Electricity Authority (NTC) uses time-series forecasting to predict peak electricity demand (e.g., 20% higher in summer due to agriculture). Their model combines:

• Historical consumption data (trend).
• Weather forecasts (seasonality).
• Economic growth indicators (causal factors). Result: Avoids blackouts by optimizing hydrothermal power plant schedules.

2. Types of Forecasting Methods

Forecasting methods are classified into two broad categories:

A. Qualitative Forecasting (Judgmental)

Used when data is scarce or expert insights are critical.

1. Delphi Method:
   • Experts anonymously share opinions iteratively until consensus.
   • Example: Nabil Bank forecasting loan demand for SMEs.
2. Market Research:
   • Surveys or focus groups (e.g., Himalayan Java testing new coffee blends).
3. Executive Judgment:
   • Top management’s intuition (e.g., Chaudhary Group expanding retail stores).

B. Quantitative Forecasting (Data-Driven)

Relies on historical data and statistical models.

1. Time-Series Models

Assumes future demand depends on past demand patterns. Decomposed into:

• Trend: Long-term increase/decrease (e.g., smartphone sales in Nepal).
• Seasonality: Repeating patterns (e.g., Daraz sales spike before Dashain).
• Cyclical: Long-term cycles (e.g., economic booms/busts).
• Random: Unpredictable fluctuations (e.g., sudden fuel shortages).

Three Key Time-Series Techniques:

Worked Example: Daraz’s Moving Average Forecast Daraz sells 500 laptops in Jan, 600 in Feb, 400 in Mar. Forecast April demand using a 3-month moving average. $${\text{MA}}_{\text{April}}=\frac{500+600+400}{3}=500\text{ laptops}$$ Limitation: Ignores seasonality (e.g., back-to-school sales in June).

2. Causal Models (Regression)

Links demand to external factors (independent variables).

• Simple Linear Regression: $Y=a+bX$ (e.g., ice cream sales vs. temperature).
• Multiple Regression: $Y=a+b_1{X}_1+b_2{X}_2$ (e.g., bank loans vs. GDP + interest rates).

Worked Example: Nabil Bank’s Loan Forecast Nabil Bank uses regression to predict loan defaults based on:

• Unemployment rate (X₁)
• Customer credit score (X₂) Model: $$\text{Default Rate}=5\%+0.3\%\times X_1-0.5\%\times X_2$$ If unemployment rises by 2% and scores drop by 10 points: $$5\%+0.6\%-0.5\%=5.1\%\text{ default rate}$$

3. Measuring Forecast Accuracy

No forecast is perfect. Error metrics quantify accuracy:

Worked Example: Pathao’s Ride Forecast Error Pathao forecasts 5,000 rides/day but gets 4,800. Calculate MAD for 5 days:

| Day | Forecast (F) | Actual (A) | Error (|F-A|) | |-----|--------------|------------|---------| | 1 | 5,000 | 4,800 | 200 | | 2 | 5,200 | 5,000 | 200 | | 3 | 4,900 | 5,100 | 200 | | 4 | 5,100 | 4,900 | 200 | | 5 | 5,000 | 5,300 | 300 | $$\text{MAD}=\frac{200+200+200+200+300}{5}=220\text{ rides}$$

4. Advanced Techniques

A. Exponential Smoothing

Adapts to recent trends using a weighting factor (α):

• α = 0.5: Equal weight to past and recent data.
• α = 0.2: Smoother, less reactive.

Worked Example: Khalti’s Transaction Forecast Khalti’s daily transactions (in millions):

B. Box-Jenkins (ARIMA)

For complex patterns (e.g., NEPSE stock prices):

• AR (Autoregressive): Uses past values.
• MA (Moving Average): Uses past forecast errors.
• I (Integrated): Accounts for trend.

In the Real World

1. NTC’s Hydroelectric Forecasting

   • Idea Used: Time-series decomposition + regression
   • How: NTC combines:
     • Historical water flow data (trend/seasonality).
     • Monsoon predictions (causal).
     • Load forecasting (exponential smoothing).
   • Impact: Avoids blackouts during peak hours (e.g., 6–9 PM).

2. Daraz’s Inventory Optimization

   • Idea Used: Moving averages + MAPE
   • How: Daraz uses a 3-month moving average for non-seasonal items (e.g., electronics) and seasonal decomposition for Dashain gifts. If MAPE > 15%, they adjust supplier orders.
   • Impact: Reduces overstocking by 20%.

3. Nabil Bank’s Loan Risk Modeling

   • Idea Used: Multiple regression
   • How: Predicts loan defaults using:
     • Unemployment rate (β₁ = +0.3).
     • Customer credit score (β₂ = –0.5).
     • Inflation rate (β₃ = +0.2).
   • Impact: Reduces bad loans by 12% via targeted credit checks.

Exam Tip

How This Unit is Tested

1. Definitions & Concepts (20%)

   • Expect questions like:
     • "Differentiate between qualitative and quantitative forecasting."
     • "What is the difference between trend and seasonality?"
   • Answer Tip: Use bullet points and real examples (e.g., "NTC uses time-series for electricity demand").

2. Calculations (30%)

   • Moving averages, exponential smoothing, MAD/MAPE.
   • Worked Example:

     Given: Demand = [100, 120, 110, 130]. Calculate 4-month MA for May. Solution: $${\text{MA}}_{\text{May}}=\frac{100+120+110+130}{4}=115$$

3. Case Studies (30%)

   • Scenario: "A local bakery sells 500 loaves in winter, 800 in summer. Forecast next winter using exponential smoothing (α=0.3)."
   • Answer Tip:
     • Identify seasonality (summer vs. winter).
     • Use decomposition or simple average if no trend.
     • Mention limitations (e.g., ignores holidays).

4. Applications (20%)

   • Question: "How would you forecast demand for a new smartphone in Nepal?"
   • Answer Structure:
     1. Qualitative: Market research (surveys in Kathmandu).
     2. Quantitative: Time-series (past smartphone sales) + regression (income levels).
     3. Error Check: Use MAPE to validate.

Common Mistakes to Avoid

• Ignoring seasonality: Always check for patterns (e.g., Daraz sales before Dashain).
• Overfitting: Using complex models (e.g., ARIMA) when simple MA works.
• No units in answers: Always label forecasts (e.g., "500 units," not just "500").
• Assuming linearity: Not all data is linear—plot a graph first!

Final Visual Summary