Learning Outcomes:
| Week | Topic | Status |
|---|---|---|
| 1 | Forecasting Foundations and Data Preparation | ✓ Complete |
| 2 | Generative AI in Forecasting and Moving Averages | ← Today |
| 3 | Stationarity, Correlation Analysis, Time Series Properties | Upcoming |
| 4 | Introductory Forecasting Methods | Upcoming |
Key Insight: Generative AI transforms forecasting from a purely statistical exercise into an intelligent, interpretive process that enhances human decision-making.
LLMs are advanced AI systems that understand, generate, and interact with human language at scale using deep learning and vast datasets.
Efficiently processes text sequences and captures complex patterns in data relationships.
Billions or trillions of parameters enable learning of grammar, semantics, and factual knowledge.
| Model | Developer | Primary Use in Forecasting |
|---|---|---|
| ChatGPT | OpenAI | Data generation, scenario analysis |
| Claude | Anthropic | Analytical reasoning, code assistance |
| Gemini | Multi-modal data integration |
LLMs identify complex temporal patterns and anomalies in high-dimensional datasets that traditional statistical methods may overlook.
Transform numerical forecasts into meaningful business narratives, explaining trends, seasonality, and confidence intervals in clear language.
Combine time series data with external text sources, news sentiment, and market indicators for comprehensive forecasting models.
Which of the following is NOT a primary advantage of using LLMs in forecasting?
Improves accuracy and adaptability through intelligent model optimization and continuous learning algorithms.
Handles vast datasets efficiently with intelligent processing pipelines that scale seamlessly with data volume.
Enables advanced "what-if" exploration through sophisticated scenario modeling and simulation capabilities.
Creates high-quality synthetic data that preserves statistical properties while ensuring privacy and compliance.
GenAI identifies statistical properties and temporal patterns in existing datasets
Creates realistic time series data maintaining original characteristics
Tests model robustness across diverse synthetic scenarios
| Challenge | Solution |
|---|---|
| Data Scarcity | Fills gaps when historical data is limited or unavailable |
| Privacy Concerns | Protects sensitive information while enabling model training |
| Development Costs | Faster and less expensive than collecting real-world data |
Key Takeaway: Synthetic data generation enables organizations to develop forecasting models 3× faster while maintaining 100% privacy compliance and achieving 50% cost reduction.
Scenario simulation transforms forecasting into a strategic tool for preparedness and resilience by exploring multiple potential futures.
Model synthetic catastrophic events (e.g., Category 5 hurricane). GenAI can:
A retail company wants to forecast demand for a new product with no historical sales data. What is the BEST use of synthetic data in this scenario?
"The quality of your AI output is directly proportional to the specificity and context of your prompts"
Provide comprehensive background about your data, industry, and forecasting objectives to guide AI reasoning.
Example: "Forecast demand for electric vehicles in Australia over the next 5 years"
Specify statistical requirements, confidence levels, and business constraints to ensure practical outputs.
Example: "Assume oil prices remain above USD $100/barrel and government subsidies continue"
Request structured responses with clear explanations, uncertainty quantification, and actionable insights.
Example: "Present results in a table with columns for scenario, assumptions, and business impact"
| Principle | Poor Example | Good Example |
|---|---|---|
| Define Context | "Forecast demand" | "Forecast quarterly demand for smartphones in Southeast Asia, 2025-2027" |
| Specify Variables | "Include important factors" | "Include GDP growth, competitor pricing, and seasonal trends" |
| Request Scenarios | "Give me a forecast" | "Provide best-case (20%), baseline (50%), and worst-case (20%) scenarios" |
| Structure Output | "Tell me the results" | "Create a table with columns: Scenario, Key Assumptions, Forecast Range, Confidence Level" |
"Generate customer data"
Problems:
"Generate 200 rows of synthetic retail customer data with variables: Customer_ID, Age (18-65), Location (Australian cities), Average_Spend, Preferred_Channel. Ensure customers under 25 spend less than $200 on average, and 30% prefer online shopping."
Strengths:
Scenario: You are forecasting weekly sales for a coffee shop chain with 15 locations across Brisbane.
Requirements:
"Generate ____ rows of synthetic weekly sales data for ____. Include variables: ____. Apply the following business rules: ____. Ensure seasonal patterns: ____."
Discussion: How would you validate that the synthetic data is realistic?
You want to create synthetic data for forecasting hospital bed occupancy. Which prompt element is MOST critical for realistic data generation?
AI models may perpetuate historical biases present in training data. Regular auditing and diverse validation datasets are essential for equitable forecasting.
Maintain clear audit trails and provide interpretable explanations for business stakeholders. Avoid "black box" decision-making.
Ensure sensitive time series data is protected when using cloud-based AI services. Consider on-premises solutions for highly confidential forecasting projects.
| Appropriate Use ✓ | Inappropriate Use ✗ |
|---|---|
| Brainstorming ideas and draft structures | Submitting AI-generated text as original work |
| Summarizing papers to aid comprehension | Using AI to avoid reading assigned texts |
| Refining or proofreading your own writing | Having AI complete entire assessments |
| Generating synthetic data for testing (if permitted) | Misrepresenting AI data as empirical research |
| Debugging code with AI assistance | Copying AI-generated code without understanding |
Always cite GenAI use: Include prompts and responses in appendices, and acknowledge AI assistance in your methodology section.
Begin with AI-assisted interpretation and code generation before implementing fully automated workflows.
AI enhances but never replaces domain expertise and critical thinking in forecasting decisions.
Stay current with emerging AI forecasting tools and techniques through hands-on experimentation.
Now that you understand how GenAI can support forecasting, we will apply these concepts to a fundamental forecasting technique: Moving Averages. You will use both traditional methods AND GenAI to deepen your understanding.
Smoothing techniques reduce random variation or noise in data to reveal underlying trends and patterns.
Smoothing filters out short-term fluctuations
↓
Highlights long-term signals
Moderate smoothing, retains more detail, responds faster to changes
Greater smoothing, clearer trend, slower response to changes
A k-period moving average uses the arithmetic mean of the k most recent time periods to forecast the next value.
| Symbol | Meaning | Example |
|---|---|---|
| Ŷt+1 | Forecast for next period | Sales forecast for Week 6 |
| Yt | Actual value at current period | Actual sales in Week 5 |
| k | Number of periods to average | 3 weeks, 4 quarters, etc. |
| t | Current time period | Week 5, Quarter 3, etc. |
Scenario: Coffee shop daily cup sales (3-day moving average)
| Day | Actual Sales | 3-Day Moving Average | Calculation |
|---|---|---|---|
| Monday | 120 | - | Not enough data |
| Tuesday | 115 | - | Not enough data |
| Wednesday | 118 | - | Not enough data |
| Thursday | 125 | 117.67 | (120 + 115 + 118) / 3 |
| Friday | 130 | 119.33 | (115 + 118 + 125) / 3 |
| Saturday | - | 124.33 | (118 + 125 + 130) / 3 ← Forecast |
Interpretation: Based on the last 3 days (Wed-Fri), we forecast Saturday sales of approximately 124 cups.
A store's daily sales for the past 4 days were: 50, 60, 55, 65 units. Using a 3-day moving average, what is the forecast for day 5?
Solution: (60 + 55 + 65) / 3 = 180 / 3 = 60 units
The choice of k (number of periods) significantly impacts forecast performance:
Advantages:
Disadvantages:
Advantages:
Disadvantages:
Practical Guideline: For seasonal data, set k equal to the season length (e.g., k=4 for quarterly data, k=12 for monthly data with yearly seasonality).
When k is even (e.g., k=4 for quarterly data), the standard moving average falls between time periods. A centered moving average aligns the average with actual time periods.
A 4-period moving average ending at period t is centered at time (t - 1.5), which is between periods. To center it at period (t - 2), we take the average of two consecutive 4-period moving averages.
For even k, the centered moving average at time t - k/2 is a weighted average:
This is equivalent to averaging two consecutive k-period moving averages.
Note: Centered moving averages cannot be used for forecasting future values, as they require future data. They are used for historical trend analysis and seasonal decomposition.
Quarterly Sales Data ($ millions):
| Quarter | Sales | 4-Period MA | Centered MA | Calculation |
|---|---|---|---|---|
| Q1 | 4.71 | - | - | Not enough data |
| Q2 | 4.75 | - | - | Not enough data |
| Q3 | 4.63 | - | 4.71 | (0.5×4.71 + 4.75 + 4.63 + 4.74 + 0.5×4.19) / 4 |
| Q4 | 4.74 | 4.71 | - | (4.71 + 4.75 + 4.63 + 4.74) / 4 |
| Q5 | 4.19 | 4.58 | - | (4.75 + 4.63 + 4.74 + 4.19) / 4 |
Interpretation: The centered moving average of 4.71 at Q3 represents the smoothed trend level for that quarter, removing seasonal fluctuations.
Monthly iPhone demand (millions of units):
| Month | 1 | 2 | 3 | 4 | 5 | 6 |
|---|---|---|---|---|---|---|
| Demand | 13 | 17 | 19 | 23 | 24 | ? |
| 2-Month MA | - | - | ? | ? | ? | ? |
Questions:
Hint: Start by calculating MA for month 3: (13 + 17) / 2 = 15
Which statement about moving averages is TRUE?
Observation: Notice how the 7-period moving average (blue) is much smoother but lags behind the original data more than the 3-period average (orange). The choice of k depends on your forecasting objectives and data characteristics.
In the practical session, you will implement moving averages using Python and pandas:
Try asking ChatGPT or Claude: "Write Python code to calculate a 4-period moving average on a DataFrame column called 'Revenue' and explain how it works."
Moving averages work best when there is no strong trend. With a clear upward or downward trend, forecasts will systematically lag behind actual values.
The first (k-1) periods cannot have a moving average calculated. For small datasets, this can be a significant limitation.
Simple moving averages treat all observations equally. Recent data may be more relevant than older data, but this method doesn't account for that.
Moving averages provide only one-step-ahead forecasts. They cannot project multiple periods into the future.
Next Steps: More sophisticated methods like Exponential Smoothing (Week 4) and ARIMA (Week 7) address these limitations.
We will explore time series properties essential for advanced forecasting methods, including stationarity tests, autocorrelation, and data transformations.