Time Series & Forecasting

Time Series Fundamentals and Patterns

  • Forecasting Assumptions and Pattern Components:

    • Time series forecasting models assume that observed data consists of a combination of systematic and unsystematic patterns.

    • Systematic Patterns: Stable, predictable structures caused by underlying physical or economic phenomena. These include persistent trends, seasonal fluctuations, and cyclical behavior.

    • Unsystematic Patterns: Unpredictable variations caused by the presence of random, irregular, or residual error terms.

  • Core Time Series Components:

    • Trend: Represents the long-term movement or baseline direction in a time series over an extended period.

    • Directionality: Indicates whether the data is generally increasing, decreasing, or remaining constant.

    • Strategic Value: Crucial for detecting structural growth or decay, allowing organizations to project the broader future trajectory of a metric.

    • Seasonality: Refers to periodic, repetitive fluctuations that recur at regular, fixed intervals within a short time frame.

    • Influencing Drivers: Governed by external calendar or temporal factors, such as climate/seasons, calendar months, or specific days of the week.

    • Operational Value: Essential for predicting short-term recurring spikes and dips, enabling tactical resource allocation and planning, particularly in seasonal industries.

    • Cyclical Behavior (Cycles):

    • Characteristics: Recurrent, wave-like oscillations that resemble seasonal patterns but operate over longer, non-strictly periodic timeframes.

    • Influencing Drivers: Closely tied to broader macroeconomic shifts, multi-year business cycles, or structural market dynamics.

    • Strategic Value: Vital for macroeconomic analysis, corporate long-term planning, and enterprise risk management.

    • Trend with a Seasonal Pattern:

    • A composite time series structure featuring a clear long-term direction (upward or downward) combined with regular short-term periodic fluctuations.

    • Data Interpretation: Short-term datasets often clearly display seasonal fluctuations and linear trends, while multi-year cyclical patterns remain difficult to identify without long-term historical records.

Moving Average Forecasting Methods

  • Methodology Overview: Moving average techniques smooth out short-term random noise by calculating average performance over a moving historical window of MM periods.

  • Simple Moving Average (SMA):

    • Concept: Takes a specific subset of the most recent MM historical periods and computes their arithmetic mean to predict the value for the upcoming period.

    • General Equation:     Y^t+1=∑i=t−M+1tYiM\hat{Y}_{t+1} = \frac{\sum_{i=t-M+1}^{t} Y_i}{M}

    • Case Application (5-Week Production Analysis):

    • Week 4 Forecast Calculation (T=4T=4):

      • Uses actual production inputs from Week 1 (Y1Y_1), Week 2 (Y2Y_2), and Week 3 (Y3Y_3).

      • Formula:         Y^4=Y1+Y2+Y33\hat{Y}_4 = \frac{Y_1 + Y_2 + Y_3}{3}

      • Numerical Value: Computes to approximately 7,5557,555.

    • Week 5 Forecast Calculation (T=5T=5):

      • Shifts the three-period window forward to include production inputs from Week 2 (Y2Y_2), Week 3 (Y3Y_3), and Week 4 (Y4Y_4).

      • Formula:         Y^5=Y2+Y3+Y43\hat{Y}_5 = \frac{Y_2 + Y_3 + Y_4}{3}

      • Numerical Value: Computes to 7,6637,663

  • Weighted Moving Average (WMA):

    • Concept: Modifies the simple moving average by assigning distinct weight ratios (wiw_i) to individual historical periods, prioritizing certain periods over others.

    • Weight Constraint Rule: To ensure unbiased estimates and prevent systematic mathematical errors, the sum of all period weights must strictly equal 1:     ∑wi=1\sum w_i = 1

Exponential Smoothing

  • Limitations of Moving Average Models:

    • Arbitrary Window Selection: Choosing the number of historical periods (MM or nn) relies on manual trial-and-error optimization, making model tuning tedious.

    • Equal Weight Allocation Flaw: Simple moving averages treat all periods within the sliding window identically, ignoring the empirical reality that recent observations generally carry higher predictive value than older observations.

  • Exponential Smoothing Foundations:

    • Utilizes all available historical data points rather than cutting off data past a arbitrary time window.

    • Applies exponentially decreasing weights to historical observations as they move further into the past.

  • Mathematical Model and Variables:

    • Forecasting Equation for Next Period (t+1t+1):     Y^t+1=αYt+(1−α)Y^t\hat{Y}_{t+1} = \alpha Y_t + (1 - \alpha) \hat{Y}_t

    • Parameter Definitions:

    • Y^t+1\hat{Y}_{t+1}: Forecast for period t+1t+1

    • YtY_t: Actual observed value in period tt

    • Y^t\hat{Y}_t: Forecasted value previously established for period tt

    • α\alpha: Smoothing constant, constrained within the continuous range:       0≤α≤10 \le \alpha \le 1

  • Behavioral Mechanics of the Smoothing Constant (Alpha):

    • Large Alpha Values (High Sensitivity):

    • Accelerates the decay rate of historical weights.

    • Places predominant weight on the most recent actual observation (YtY_t).

    • Enhances responsiveness to sudden, recent changes in data.

    • Small Alpha Values (High Smoothing):

    • Slows the decay rate of past observations.

    • Retains significant weight on older historical context and past predictions (Y^t\hat{Y}_t).

    • Effectively filters out temporary fluctuations and short-term noise.

  • Step-by-Step Exponential Smoothing Example (Alpha = 0.2):

    • Week 1:

    • Forecast cannot be mathematically computed due to the absence of prior observational data.

    • Week 2 (Initialization Process):

    • System initialization sets the Week 2 forecast directly equal to the actual value observed in Week 1:       Y^2=Y1\hat{Y}_2 = Y_1

    • Week 3 Calculation:

    • Combines actual Week 2 observation (Y2Y_2) and Week 2 forecast (Y^2\hat{Y}_2):       Y^3=0.2Y2+(1−0.2)Y^2\hat{Y}_3 = 0.2 Y_2 + (1 - 0.2) \hat{Y}_2

    • Numerical Value: Computes to 7,647.67,647.6

    • Week 4 Calculation:

    • Combines actual Week 3 observation (Y3Y_3) and Week 3 forecast (Y^3=7,647.6\hat{Y}_3 = 7,647.6):       Y^4=0.2Y3+(1−0.2)Y^3\hat{Y}_4 = 0.2 Y_3 + (1 - 0.2) \hat{Y}_3

    • Numerical Value: Computes to 7,541.487,541.48