Comprehensive Study Notes on Time Series, Causal Forecasting, and Forecast Error Metrics

Time Series Forecasting Models

  • Core Concepts of Time Series Models:

    • Time series forecasting models rely strictly on chronological historical data where time acts as the primary independent variable.

    • Exam problems feature distinct numerical datasets requiring systematic application of moving average, weighted moving average, and exponential smoothing formulas.

  • Simple Three-Period Moving Average:

    • Definition: Calculates the arithmetic mean of actual demand from the most recent three periods to forecast the next period.

    • Weight Distribution: Assigns equal weight to each historical period (13\frac{1}{3} or 3313%33\frac{1}{3}\% per period).

    • Moving Mechanism: Drops the oldest actual demand figure and incorporates the newest actual demand figure as time advances.

    • Governing Formula: Ft=Dt−1+Dt−2+Dt−33F_t = \frac{D_{t-1} + D_{t-2} + D_{t-3}}{3}

    • Historical Dataset:

      • Period 10 actual demand: 2.482.48

      • Period 11 actual demand: 3.703.70

      • Period 12 actual demand: 4.244.24

      • Period 13 actual demand: 2.862.86

      • Period 14 actual demand: 4.444.44

    • Step-by-Step Calculations:

      • Forecast for Period 13 (F13F_{13}):

        • Input data: Actual demand from periods 10, 11, and 12 (2.482.48, 3.703.70, 4.244.24).

        • Calculation: F13=2.48+3.70+4.243=10.423=3.47F_{13} = \frac{2.48 + 3.70 + 4.24}{3} = \frac{10.42}{3} = 3.47

      • Forecast for Period 14 (F14F_{14}):

        • Input data: Actual demand from periods 11, 12, and 13 (3.703.70, 4.244.24, 2.862.86).

        • Calculation: F14=3.70+4.24+2.863=10.803=3.60F_{14} = \frac{3.70 + 4.24 + 2.86}{3} = \frac{10.80}{3} = 3.60

      • Forecast for Period 15 (F15F_{15}):

        • Input data: Actual demand from periods 12, 13, and 14 (4.244.24, 2.862.86, 4.444.44).

        • Calculation: F15=4.24+2.86+4.443=11.543=3.85F_{15} = \frac{4.24 + 2.86 + 4.44}{3} = \frac{11.54}{3} = 3.85

  • Two-Period Weighted Moving Average:

    • Definition: Assigns specific, unequal weighting factors to prior demand observations, placing higher value on more recent data.

    • Weight Allocation Rules:

      • Most recent observation weight: 0.700.70 (70%70\%).

      • Older observation weight: 0.300.30 (30%30\%).

      • Sum of all weights must equal 1.001.00 (100%100\%).

    • Governing Formula: Ft=(0.70×Dt−1)+(0.30×Dt−2)F_t = (0.70 \times D_{t-1}) + (0.30 \times D_{t-2})

    • Step-by-Step Calculations:

      • Forecast for Period 12 (F12F_{12}):

        • Input data: Period 11 demand (3.703.70) weighted at 70%70\%, Period 10 demand (2.482.48) weighted at 30%30\%.

        • Calculation: F12=(0.70×3.70)+(0.30×2.48)=2.59+0.744=3.334≈3.33F_{12} = (0.70 \times 3.70) + (0.30 \times 2.48) = 2.59 + 0.744 = 3.334 \approx 3.33

      • Forecast for Period 13 (F13F_{13}):

        • Input data: Period 12 demand (4.244.24) weighted at 70%70\%, Period 11 demand (3.703.70) weighted at 30%30\%.

        • Calculation: F13=(0.70×4.24)+(0.30×3.70)=2.968+1.11=4.078≈4.08F_{13} = (0.70 \times 4.24) + (0.30 \times 3.70) = 2.968 + 1.11 = 4.078 \approx 4.08

      • Forecast for Period 14 (F14F_{14}):

        • Input data: Period 13 demand (2.862.86) weighted at 70%70\%, Period 12 demand (4.244.24) weighted at 30%30\%.

        • Calculation: F14=(0.70×2.86)+(0.30×4.24)=2.002+1.272=3.274≈3.27F_{14} = (0.70 \times 2.86) + (0.30 \times 4.24) = 2.002 + 1.272 = 3.274 \approx 3.27

      • Forecast for Period 15 (F15F_{15}):

        • Input data: Period 14 demand (4.444.44) weighted at 70%70\%, Period 13 demand (2.862.86) weighted at 30%30\%.

        • Calculation: F15=(0.70×4.44)+(0.30×2.86)=3.108+0.858=3.966≈3.97F_{15} = (0.70 \times 4.44) + (0.30 \times 2.86) = 3.108 + 0.858 = 3.966 \approx 3.97

Exponential Smoothing and Alpha Parameter Selection

  • Exponential Smoothing Fundamentals:

    • Definition: A sophisticated time series technique that weights past observations with exponentially decreasing weights, balancing immediate actual demand against prior forecasted values.

    • Governing Formula: Ft+1=α×Dt+(1−α)×FtF_{t+1} = \alpha \times D_t + (1 - \alpha) \times F_t

      • α\alpha = Alpha parameter (smoothing constant), where 0≤α≤10 \le \alpha \le 1

      • DtD_t = Actual demand in period tt

      • FtF_t = Forecasted demand for period tt

    • Exam Memorization Rule: Time series formulas (moving average, weighted moving average, exponential smoothing) will not be provided on exams and must be completely memorized.

  • Step-by-Step Calculation Walkthrough (α=0.7\alpha = 0.7):

    • Week 2 Forecast (F2F_2):

      • Given baseline: Week 1 actual demand (D1D_1) = 1010, Week 1 initial forecast (F1F_1) = 1010.

      • Calculation: F2=(0.7×10)+(1−0.7)×10=(0.7×10)+(0.3×10)=7.0+3.0=10.0F_2 = (0.7 \times 10) + (1 - 0.7) \times 10 = (0.7 \times 10) + (0.3 \times 10) = 7.0 + 3.0 = 10.0

    • Week 3 Forecast (F3F_3):

      • Given data: Week 2 actual demand (D2D_2) = 1111, Week 2 forecast (F2F_2) = 10.010.0

      • Calculation: F3=(0.7×11)+(1−0.7)×10.0=7.7+3.0=10.7F_3 = (0.7 \times 11) + (1 - 0.7) \times 10.0 = 7.7 + 3.0 = 10.7

    • Week 4 Forecast (F4F_4):

      • Given data: Week 3 actual demand (D3D_3) = 99, Week 3 forecast (F3F_3) = 10.710.7

      • Calculation: F4=(0.7×9)+(1−0.7)×10.7=6.3+3.21=9.51F_4 = (0.7 \times 9) + (1 - 0.7) \times 10.7 = 6.3 + 3.21 = 9.51

  • Alpha Parameter Selection Rules for Random Data:

    • High Alpha Value (e.g., α=0.8\alpha = 0.8):

      • Places 80%80\% weight on immediate actual demand (DtD_t) and 20%20\% weight on past forecast (FtF_t).

      • Consequence with Random Data: Causes the forecast to heavily chase random spikes and noise. The forecast curve moves out of phase, lagging behind demand (e.g., raising forecast immediately after demand peaked, lowering forecast right after demand dropped).

    • Low Alpha Value (e.g., α=0.2\alpha = 0.2):

      • Places 20%20\% weight on immediate actual demand (DtD_t) and 80%80\% weight on past forecast (FtF_t).

      • Consequence with Random Data: Effectively smooths out random noise, providing a stable trendline that reduces cumulative average forecast error.

Linear Regression and Demand Patterns

  • Linear Regression Overview:

    • Definition: A quantitative time series model fitting a linear trendline to historical chronological data.

    • Algebraic Form: y=mx+by = mx + b

      • yy = Dependent variable (forecasted value to be solved).

      • xx = Independent variable (chronological time period / month).

      • mm = Slope coefficient (rate of change per time unit).

      • bb = Intercept term (y-axis crossing point).

    • Exam Policy: Coefficients (mm) and intercepts (bb) will be supplied in exam prompts; manual computation of regression constants is not required.

  • Clem's Drag Racing Competition Clutches Example:

    • Dataset context: Historical sales gathered across months 1 through 10.

    • Provided Regression Equation: y=4.49+7.23×xy = 4.49 + 7.23 \times x

    • Month 11 Forecast (F11F_{11}):

      • Calculation: y=4.49+(7.23×11)=4.49+79.53=84.02y = 4.49 + (7.23 \times 11) = 4.49 + 79.53 = 84.02 (or 85.1285.12 using 7.337.33 coefficient variation).

    • Month 12 Forecast (F12F_{12}):

      • Calculation: y=4.49+(7.23×12)=4.49+86.76=91.25y = 4.49 + (7.23 \times 12) = 4.49 + 86.76 = 91.25 (or 96.496.4).

    • Monthly Incremental Rate: Demand grows linearly by 7.237.23 clutches per month.

  • Four Primary Demand Patterns:

    1. Trend: Overall upward or downward directional movement over extended timeframes.

    2. Seasonality: Recurring, space-and-time-bound demand spikes and dips within annual calendar periods (e.g., drag racing parts demand dropping in Pennsylvania during January and peaking in June).

    3. Cyclicality: Multi-year recurring wave patterns (typically spanning 3 to 5 years).

    4. Randomness: Unpredictable, unsystematic variations in demand.

  • Limitations of Linear Regression:

    • Assumes a strict, endless straight-line trend.

    • Fails to capture abrupt market changes or new market entrants (e.g., a rival opening Mary's Competition Clutch Shop).

    • Ignores seasonality unless modified by seasonal indices or adjustments.

    • Industry applications requiring seasonal adjustments: Gasoline usage, children's apparel, medical services, and vacation travel.

Causal Forecasting and Multiple Regression

  • Causal Forecasting Principles:

    • Definition: Shift from time-dependent models to cause-and-effect models where external drivers dictate demand shifts.

    • Real-World Application Examples:

      • NGO Drought Relief Expenditures: Driven by rainfall volume (inches of precipitation), not calendar dates.

      • Financial Mortgage Applications: Driven by prevailing mortgage interest rates. When mortgage rates surge, applications plunge; when rates plummet, refinancing applications surge.

  • Multiple Regression Model Structure:

    • Governing Formula: y=a+(b1×x1)+(b2×x2)+⋯+(bn×xn)y = a + (b_1 \times x_1) + (b_2 \times x_2) + \dots + (b_n \times x_n)

      • yy = Dependent variable (forecasted demand).

      • aa = Intercept parameter.

      • bib_i = Weighting coefficients for specific driver variables.

      • xix_i = Independent causal variables (x1,x2,…,xnx_1, x_2, \dots, x_n).

  • Qualitative vs. Quantitative Method Selection:

    • Quantitative Forecasting: Employed when extensive historical data exists (e.g., Hershey Kisses possessing 120 to 130 years of historical brand data).

    • Qualitative Forecasting: Employed for startups or totally novel product launches where historical sales data is zero.

  • Causal Variables in Confectionery Forecasting (Hershey Example):

    • Seasonal holiday calendars.

    • Internal advertising expenditures.

    • Competitor spend, competitive entries/exits, and rival product launches.

    • Pricing adjustments (price increases generally cause inverse drops in demand).

    • Macro confectionery trends (e.g., shifting health and dietary preferences).

Causal Model Application: TJ's Ice Cream Example

  • Scenario Context:

    • TJ forecasts weekly ice cream production (in gallons) using weekly high temperatures and advertising spend (analogous to forecasting Burt's Bees lip balm sales via wind chill indices and ad budgets at Clorox).

  • Model Parameters and Inputs:

    • Dependent Variable (yy): Gallons of ice cream to produce.

    • Independent Variable 1 (x1x_1): Average weekly high temperature (74∘F74^\circ\text{F}).

    • Independent Variable 2 (x2x_2): Weekly advertising spend (475\text{475}).

    • Calculated Intercept (aa): −438-438

    • Temperature Coefficient (b1b_1): 14.8914.89

    • Advertising Coefficient (b2b_2): 0.130.13

    • Provided Regression Equation: y=−438+(14.89×x1)+(0.13×x2)y = -438 + (14.89 \times x_1) + (0.13 \times x_2)

  • Step-by-Step Calculation:

    • Temperature component: 14.89×74=1101.8614.89 \times 74 = 1101.86

    • Advertising component: 0.13×475=61.750.13 \times 475 = 61.75

    • Intercept component: −438-438

    • Total Forecast (yy): y=−438+1101.86+61.75=725.61≈725 gallonsy = -438 + 1101.86 + 61.75 = 725.61 \approx 725\,\text{gallons} (or 724.98 gallons724.98\,\text{gallons}).

  • Interpretation of Model Intercept:

    • The negative intercept (−438-438) is a mathematical line-fitting constant. It does not literally mean that at 0∘F0^\circ\text{F} and $0\$0 advertising spend the company owes customers 438 gallons of ice cream.

  • First Law of Forecasting:

    • Statement: "Forecasts are almost always wrong, but useful."

    • Even with inevitable error, generating a 725-gallon forecast prevents huge mistakes, such as overproducing during cold 50∘F50^\circ\text{F} weeks or underproducing during extreme heat waves.

Forecast Error Measurement and Model Accuracy

  • Fundamental Error Formula:

    • Forecast Errort=Dt−Ft\text{Forecast Error}_t = D_t - F_t

    • Order rule: Actual demand (DtD_t) minus forecasted demand (FtF_t) ("DD comes before FF in the alphabet").

  • Key Forecast Error Metrics:

    1. Mean Forecast Error (MFE):

      • Formula: MFE=∑(Dt−Ft)n\text{MFE} = \frac{\sum (D_t - F_t)}{n}

      • Measures central tendency and directional forecast bias.

      • Flaw: Positive errors (+10+10) and negative errors (−10-10) cancel out to zero (MFE=0\text{MFE} = 0), masking severe forecast inaccuracies.

    2. Mean Absolute Deviation (MAD):

      • Formula: MAD=∑∣Dt−Ft∣n\text{MAD} = \frac{\sum |D_t - F_t|}{n}

      • Measures overall error magnitude, treating over-forecasting and under-forecasting as equally harmful.

    3. Mean Absolute Percentage Error (MAPE):

      • Formula: MAPE=∑(∣Dt−Ft∣Dt)n×100%\text{MAPE} = \frac{\sum \left( \frac{|D_t - F_t|}{D_t} \right)}{n} \times 100\%

      • Preferred metric because it evaluates error relative to total demand scale, permitting direct comparison across high-volume and low-volume products.

  • Volume Scaling and Metric Comparison Example (Item A vs. Item B):

    • Item A: MAD=100,000 units\text{MAD} = 100{,}000\,\text{units}, Average Monthly Volume = 500,000 units500{,}000\,\text{units}.

      • MAPE=100,000500,000=20%\text{MAPE} = \frac{100{,}000}{500{,}000} = 20\%

    • Item B: MAD=200,000 units\text{MAD} = 200{,}000\,\text{units}, Average Monthly Volume = 4,000,000 units4{,}000{,}000\,\text{units}.

      • MAPE=200,0004,000,000=5%\text{MAPE} = \frac{200{,}000}{4{,}000{,}000} = 5\%

    • Evaluation: Item B possesses the superior forecast despite having double the absolute deviation (200,000200{,}000 vs 100,000100{,}000) because its percentage error is substantially lower (5%5\% vs 20%20\%).

  • Detailed Step-by-Step Model Comparison Walkthrough (Model 1):

    • Period 1: Demand D1=18D_1 = 18, Forecast F1=20F_1 = 20

      • Error: 18−20=−218 - 20 = -2

      • Absolute Deviation: ∣−2∣=2|-2| = 2

      • Absolute Percent Error: \frac{2}{18} = 11.11\%

    • Period 2: Demand D2=14D_2 = 14, Forecast F2=18F_2 = 18

      • Error: 14−18=−414 - 18 = -4

      • Absolute Deviation: ∣−4∣=4|-4| = 4

      • Absolute Percent Error: \frac{4}{14} = 28.57\%

    • Period 3: Demand D3=21D_3 = 21, Forecast F3=19F_3 = 19

      • Error: 21−19=221 - 19 = 2

      • Absolute Deviation: ∣2∣=2|2| = 2

      • Absolute Percent Error: \frac{2}{21} = 9.52\%

    • Model 1 Summary Averages: MFE=0.7\text{MFE} = 0.7, MAD=4.1\text{MAD} = 4.1, MAPE=21.26%\text{MAPE} = 21.26\%

    • Model Selection: Model 2 is chosen over Model 1 due to lower MFE, lower MAD, and lower MAPE.

Forecast Bias and Operational Impacts

  • Over-Forecasting (Negative MFE):

    • Occurs when forecast consistently exceeds demand (Ft>DtF_t > D_t, yielding Dt−Ft<0D_t - F_t < 0).

    • Operational Consequences: Drives unnecessary manufacturing, builds excess inventory, spikes warehouse holding costs, and increases product obsolescence.

    • Organizational Root Cause: "The Beer Game" behavioral effect—sales and marketing personnel distrust supply chain delivery reliability, prompting them to deliberately inflate forecast orders.

  • Under-Forecasting (Positive MFE):

    • Occurs when forecast falls below actual demand (Ft<DtF_t < D_t, yielding Dt−Ft>0D_t - F_t > 0).

    • Operational Consequences: Triggers product scarcity, stockouts, and lost sales revenue.

    • Financial & Production Consequences: Unplanned demand forces emergency weekend production runs. Emergency shifts require overtime premium labor rates (time-and-a-half pay). Because cost of goods sold (COGS) spikes while product sales price remains fixed, the business suffers margin compression.

  • Operational Remediation Strategy:

    • Organizations must audit and correct organizational forecast bias (trust issues, behavioral game-playing) before tweaking complex quantitative models.

Tracking Signals and Modern Automated Forecasting

  • Tracking Signal Functionality:

    • Definition: An analytical control mechanism used to track whether forecast errors stay within acceptable control limits over time.

    • Operational Mechanism: Tracks error boundaries (e.g., normal variation band of ±10%\pm 10\%). If cumulative error drifts out to ±20%\pm 20\% or ±30%\pm 30\% absolute deviation, the tracking signal trips an alert indicating structural shifts in demand.

    • Exam Policy: Manual tracking signal calculations are not required.

  • Modern Systems and Artificial Intelligence (AI):

    • Advanced enterprise systems evaluate dozens of causal drivers simultaneously.

    • Garbage In, Garbage Out: Automated and AI forecasting models rely strictly on input data quality. Flawed data inputs generate misleading output predictions.

    • Managerial Imperative: Managers must understand underlying mathematical mechanics to critically evaluate automated forecasting outputs and catch logical errors.

Questions and Classroom Discussion

  • Exam Formula Policy Prompt:

    • Question: Are students provided with formula reference sheets on exams?

    • Response: Basic time series formulas (simple moving average, weighted moving average, exponential smoothing) will NOT be provided on exams and must be memorized. Linear and multiple regression equations will be directly provided within the exam prompts.

  • Random Data Smoothing Parameter Prompt:

    • Question: When dealing with highly random demand, should a forecaster select a high or low alpha parameter?

    • Response: Select a low alpha parameter (α=0.2\alpha = 0.2). High alphas place excessive weight on random demand spikes, causing the model to chase noise and lag behind trends, whereas low alphas effectively smooth out variance.

  • Regression Coefficients Calculation Prompt:

    • Question: Will students have to compute regression intercepts and slope coefficients manually on tests?

    • Response: No. Calculating linear regression coefficients manually is too time-consuming for exam constraints; complete regression equations will be supplied directly.

  • Absolute Deviation vs. Volume Scaling Prompt:

    • Question: Is a planner with an item MAD of 100,000 units100{,}000\,\text{units} automatically better than a planner with an item MAD of 200,000 units200{,}000\,\text{units}?

    • Response: No. Evaluation depends on sales volume scale. If the 100,000 unit100{,}000\,\text{unit} MAD corresponds to a 500,000 unit500{,}000\,\text{unit} demand (20%20\% MAPE) and the 200,000 unit200{,}000\,\text{unit} MAD corresponds to a 4,000,000 unit4{,}000{,}000\,\text{unit} demand (5%5\% MAPE), the forecaster with the 200,000 unit200{,}000\,\text{unit} MAD is significantly more accurate relative to overall scale.