Forecasting: Fundamentals, Models, and Accuracy Measures

Fundamental Concepts of Forecasting

• Definition: Forecasting is both an art and a science focused on predicting future events or trends. It follows a systematic process involving the analysis of historical data to identify patterns and uses statistical models to project those patterns into future timeframes.

• Purpose: The primary goal is to provide the best possible estimate for defined timeframes to inform strategic planning and operational decisions.

• Domains of Use: It is widely applied in finance, economics, weather prediction, environmental science, and supply chain management.

• Importance of Forecasting:

  • Market Adaptation: It allows businesses to anticipate market trends, changes in consumer behavior, and shifts in economic conditions.

  • Environmental Prediction: Helps in forecasting extreme weather patterns and the impacts of climate change.

  • Financial Guidance: It is essential for investment decisions and assessing risk levels.

  • Strategy Implementation: Different departments rely on it for specific needs:

    • Finance Department: Uses it for projecting cash flows and determining capital requirements.

    • Human Resources: Predicts staffing requirements.

    • Production Department: Organizes production schedules, labor needs, raw materials, and inventory management.

    • University Context: Admissions (enrollment predictions), Finance (budgeting), Facilities Management (maintenance/development), and Academic Departments (course demand).

• The Inherent Nature of Forecasts: Forecasts are inherently imperfect. However, well-informed guesses are significantly more valuable for planning than having no forecasts at all. Organizations without forecasting struggle to adapt to change or mitigate threats.

Sector-Specific Examples of Forecasting Applications

• Environmental Conservation: Used to predict human impact on wildlife. For example, if habitat loss and climate data suggest a species decline, conservationists can implement habitat protection measures in advance.

• Weather Forecasting: Meteorologists predict temperature, precipitation, and wind speeds. Accurate forecasts allow farmers to plan irrigation, municipalities to prepare for floods, and individuals to plan activities, thereby ensuring safety and minimizing damage.

• Financial Forecasting: Analysts use historical data and market trends to predict conditions like stock market movements. For instance, predicting a rise in renewable energy stock prices due to clean energy demand allows investors to make informed buy/sell decisions.

• Supply Chain Forecasting: Predicts product demand to manage inventory and logistics. If a surge in demand for electric vehicles is forecasted, manufacturers can increase production and suppliers can secure raw materials before the surge occurs.

Decision-Making, Planning, and Urban Development

• Internal vs. External Events:

  • Internal Events: Operations, product development, and marketing campaigns which are within a company's direct control (agency).

  • External Events: Global economic shifts, consumer behavior, and regulatory changes which are beyond direct control and require forecasting to navigate.

• Conceptual Framework:

  • Decision-Making: The act of making choices within the company's control.

  • Forecasting: Predicting the external market and environmental changes.

  • Planning: The overarching blueprint that aligns internal capabilities with the external environment to achieve strategic goals.

• Example: Urban Development Planning:

  • A city council foresees population growth based on birth rates and immigration.

  • The council decides how to accommodate this growth (expanding zones vs. new developments).

  • The final plan integrates these decisions with timelines, budgets, and infrastructure upgrades needed to meet the forecasted increase.

Robust Areas and Industry Applications

• Robust Categories:

  • Scheduling: Estimating time for project completion (e.g., construction phases).

  • Acquiring Resources: Anticipating the need for new personnel (e.g., tech companies recruiting software engineers early).

  • Determining Requirements: Deciding on long-term resource needs (e.g., manufacturing firms ordering raw materials for the upcoming year based on demand targets).

• Industry Applications:

  • Financial Sector: Predicting market trends and assessing loan default risks.

  • Supply Chain: Optimizing stock levels to avoid overstocking or stockouts.

  • Healthcare: Managing staff schedules and anticipating patient influx during flu seasons.

  • Energy Sector: Predicting peak electricity usage times to adjust generation and grid operations.

  • Public Policy: Planning for infrastructure, housing, and schools based on population growth.

Analysis of Data Patterns and Trends

• Visualizing Forecasting Variables:

  • Upward Trend: Observed in datasets like Australian Monthly Electricity Production, driven by economic and population growth.

  • Volatile Trend: Observed in U.S. Treasury Bill Contracts, characterized by peaks and troughs from interest rate changes and investor sentiment.

  • Seasonal Trend: Observed in Product sales where sales rise and fall at specific times of the year due to consumer habits.

  • Cyclical Trend: Observed in Australian Clay Brick Production, reflecting construction industry demand cycles.

• Specific Visualization Examples:

  • Stock Market: Line graphs showing closing prices over time with $X$-axis representing time and $Y$-axis representing price.

  • Retail Sales: Units sold each day, highlighting promotional impacts.

  • Unemployment Rate: Quarterly state-level data used for policy-making.

  • Gasoline Prices: Daily average prices used to inform transportation cost decisions.

Quantitative, Qualitative, and Unpredictable Methods

• Quantitative Methods: Rely on numerical data and mathematical models.

  • Time Series Forecasting: Assumes future behavior is inferred from past behavior (e.g., using past retail growth to predict future sales).

  • Explanatory Modeling: Seeks to understand how variables like price or advertising spend influence a target (e.g., using regression to see how marketing affects product sales).

• Qualitative Methods: Used when data is limited or unreliable. Depends on subjective judgment and expertise.

  • Delphi Method: A structured communication technique using panels of experts.

  • Expert Surveys and Scenario Analysis: Gathering insights from professionals and constructing narratives.

• Unpredictable Methods: Used for highly uncertain or rare events.

  • Scenario Planning: Building multiple plausible futures. Example: envisioning scenarios for interplanetary travel results.

  • Wild Cards Analysis: Identifying low-probability, high-impact events like the discovery of a non-polluting, cheap energy source.

  • Sensitivity Analysis: Examining how changes in one variable impact a forecast. Example: Analyzing how a large oil price increase affects consumption.

Advanced Quantitative Models

• Machine Learning (ML) Forecasting: Uses techniques like neural networks or random forests.

  • Example: Recurrent Neural Networks (RNN) trained on historical prices and trading volume to predict stock movements.

• Ensemble Forecasting: Combines multiple models to enhance accuracy.

  • Example: Combining different meteorological models to predict hurricane landfalls.

• Bayesian Forecasting: Uses Bayesian statistics to update beliefs as new data arrives.

  • Example: Updating hospital readmission rates by incorporating demographic shifts and new treatment outcomes.

Explanatory and Time Series Models

• Explanatory Models: Focus on cause-and-effect mechanisms. They assume the variable to be forecasted ($Y$) has a relationship with independent variables ($X$).

  • Formula Example (GNP):     $GNP = f(\text{monetary policies, fiscal policies, inflation, capital spending, imports, exports, error})$

  • Key features: Variable identification, relationship establishment (linear/nonlinear), mechanism explanation, and predictive capability.

  • Epidemiological Application: Using Compartmental models like the SIR model to simulate the spread of infectious diseases (COVID-19, HIV/AIDS).

• Time Series Models: Treat systems as a "black box." They do not look for causality but focus entirely on patterns in historical data.

  • Pattern Discovery: Identifies trends (long-term), seasonality (repeating), cyclicality (regular fluctuations), and irregularities (noise).

  • Techniques: Autoregressive Integrated Moving Average (ARIMA), Holt-Winters exponential smoothing, and LSTM (Long Short-Term Memory) networks.

The Five Steps of the Forecasting Task

1. Problem Definition: Understanding the purpose, the stakeholders, and how the forecast will be used.

   • Bakery Example: Predicting daily loaf sales to minimize waste.

2. Gathering Information: Collecting statistical data and expert judgment.

   • Bakery Example: Reviewing two years of sales data and noting farmers' market days.

3. Preliminary (Exploratory) Analysis: Visualizing data to find trends, cycles, and outliers.

   • Bakery Example: Observing a spike in whole wheat bread demand in January (New Year's resolutions).

4. Choosing and Fitting Models: Selecting models (like exponential smoothing for seasonality) and adjusting parameters to fit historical data.

   • Retail Example: Fitting a model that accounts for promotional spend and economic indicators.

5. Using and Evaluating a Model: Generating the forecast and comparing predicted results to actual outcomes to assess accuracy.

   • Retail Example: Planning promotions if a downturn is predicted or optimizing inventory for a forecasted surge.

Statistical Measures in Forecasting

• Univariate Statistics (Single Variable Analysis):

  • Central Tendency: Mean ($\text{average}$), Median ($\text{middle value}$), and Mode ($\text{most frequent}$).

  • Measures of Dispersion:

    • Range: (Highest - Lowest).

    • Variance: Average of squared differences from the mean.

    • Standard Deviation: Square root of variance ($\approx 7.5$ in the classroom test example).

  • Class Score Example: Scores: $78, 85, 92, 67, 75, 80, 84, 90, 77, 82$.

    • $\text{Mean} = 81$

    • $\text{Median} = 80.5$

    • $\text{Range} = 92 - 67 = 25$

• Bivariate Statistics (Relationship between two variables $X$ and $Y$):

  • Example: Student Height ($X$) vs. Grade ($Y$). Correlation is identified via scatter plots.

• Autocorrelation (Serial Correlation): Measures similarity between a time series and a lagged version of itself over time.

  • Formula for Autocorrelation Function ($\rho(k)$):     $\rho(k) = \frac{\sum_{t=k+1}^{n} (X_t - \bar{X})(X_{t-k} - \bar{X})}{\sum_{t=1}^{n} (X_t - \bar{X})^2}$

  • Values: If $r=1$, proceed to forecasting; if $r=0$, no forecast model is needed.

Evaluating Forecast Accuracy

• Definitions:

  • Observation ($Y_t$): Actual measured value at time $t$.

  • Forecast ($F_t$): Predicted value at time $t$.

  • Error ($e_t$): $e_t = Y_t - F_t$

• Standard Metrics:

  • Mean Error (ME): Average of differences. Indicates bias ($+\text{underestimates}$, $-\text{overestimates}$).     $ME = \frac{1}{n} \sum_{i=1}^{n} e_t$

  • Mean Absolute Error (MAE/MAD): Indicates magnitude of errors regardless of direction.     $MAE = \frac{1}{n} \sum_{i=1}^{n} |e_t|$

  • Mean Squared Error (MSE): Weights larger errors more heavily.     $MSE = \frac{1}{n} \sum_{i=1}^{n} e_t^2$

  • Percentage Error (PE): $PE_t = \frac{e_t}{Y_t} \times 100$

  • Mean Percentage Error (MPE): Average percentage bias.     $MPE = \frac{1}{n} \sum_{i=1}^{n} PE_t$

  • Mean Absolute Percentage Error (MAPE): Average relative accuracy.     $MAPE = \frac{1}{n} \sum_{i=1}^{n} |PE_t|$

Advanced Regression and Naïve Models

• Method of Least Squares: Standard approach to find the best-fitting curve by minimizing the sum of squared residuals.

• Gauss-Markov Theorem: Under specific assumptions, Ordinary Least Squares (OLS) is the Best Linear Unbiased Estimator (BLUE).

• Classical Assumptions for BLUE:

  1. Linear relationship.

  2. Independent residuals.

  3. Homoscedasticity (constant variance).

  4. No perfect multicollinearity.

  5. Zero mean of residuals.

• Naïve Forecasting Methods:

  • Simple Naïve: $F_{t+1} = Y_t$. Example: Selling 120 umbrellas in Feb implies sales of 120 in March.

  • Seasonal Naïve: Forecast is the same as the same period last year. Example: 200 guests at a resort every December.

  • Moving Average Naïve: Average of a fixed number of recent observations. Example: If visitor counts are 1200, 1500, and 1300 for 3 months, Feb forecast is $\frac{1200+1500+1300}{3} = 1333.33$.

Theil’s U-Statistic and Prediction Intervals

• Theil’s U-Statistic: Compares a model to a naïve forecast while emphasizing large errors via squaring.

  • Formula:     $U = \sqrt{\frac{\sum_{t=1}^{n-1} (FPE_{t+1} - APE_{t+1})^2}{\sum_{t=1}^{n-1} (APE_{t+1})^2}}$

  • Where relative changes are:     $FPE_{t+1} = \frac{F_{t+1} - Y_t}{Y_t}$     $APE_{t+1} = \frac{Y_{t+1} - Y_t}{Y_t}$

  • Interpretation:

    1. $U = 1$: Forecasting method is no better than naïve approach.

    2. U < 1: Forecasting method is superior to naïve approach.

    3. U > 1: Better to stick with the naïve approach.

• Prediction Intervals: Uncertainty statements surrounding forecast values.

  • Formula: $F_{n+1} \pm Z_{\alpha} \sqrt{MSE}$

  • For a 95% interval, $Z = 1.96$.

Statistical Hypothesis Testing and Error Types

• Hypotheses:

  • Null Hypothesis ($H_0$): No significant difference or effect ($H_0: \mu_A = \mu_B$).

  • Alternative Hypothesis ($H_1$ or $H_a$): Contradicts $H_0$ ($H_1: \mu_A \neq \mu_B$).

• Measuring Risk:

  • Type I Error ($\alpha$): Producer's Risk. Rejecting a true $H_0$ (False Positive). Common level is $0.05$.

  • Type II Error ($\beta$): Consumer's Risk. Failing to reject a false $H_0$ (False Negative).

  • Power of the Test ($1 - \beta$): Probability of correctly rejecting a false null hypothesis.

• Decision Matrix: If the p-value $< \alpha$, reject $H_0$. Balance the risks based on cost: if false positives are expensive, set a lower $\alpha$; if false negatives are risky (e.g., life-saving drug), increase the sample size to reduce $\beta$.