Forecasting fundamentals and time series techniques (Module summary)

Forecasting fundamentals

  • A forecast is a statement about the future value of a variable of interest.
  • Forecasts are used to balance supply and demand by providing a planning tool or model; without forecasts, decisions would be guesses and less effective.
  • Two important aspects of forecasts:
    • Expected level of demand, based on structural variation such as trend or seasonal variation. Examples:
    • Web traffic for a retail store tends to trend upward as more people shop online.
    • There is seasonal variation, e.g., more hits in December than in other months.
    • Forecast accuracy, related to the potential size of forecast error; it measures how close the forecast is to the actual value.
  • Forecasting techniques share common features:
    • They assume an underlying causal system that existed in the past and will persist into the future.
    • Good managers should not rely solely on the model; they must monitor the environment for changes (e.g., tax cuts, new inventions, weather events) that can alter demand.
    • Forecasts are imperfect; randomness can interfere with accuracy.
    • Forecasts for groups of items are usually more accurate than for individual items due to the canceling effect.
  • Canceling effect (group forecasting): forecasting for a group (e.g., all trail bikes) tends to be more accurate than forecasting for individual items because random errors can cancel out when aggregated.
  • Forecast accuracy typically declines as the forecasting horizon increases: longer horizons are harder to forecast accurately (e.g., 20-year vs 10-year vs 1-year forecasts).
  • Good forecast attributes:
    • Timely: provides enough time to ramp up production if needed.
    • Accurate: has a measurable error bound.
    • Reliable: stable across runs; avoid garbage-in garbage-out.
    • Expressed in meaningful units (e.g., dollars, units, etc.) and in a usable format for the audience.
    • Simple to understand: not overly complicated; stakeholders should have confidence in the forecast.
  • Simpler forecasting techniques (e.g., naive approach) are popular because they are easy to understand, even if they may be less accurate.

Forecasting process steps

  • Step 1: Determine the purpose of the forecast, which informs the required level of detail (e.g., big-picture planning vs. production scheduling for the next twelve months).

  • Step 2: Establish the time horizon (daily, monthly, yearly) to understand the required accuracy and suitable techniques.

  • Step 3: Gather historical data; obtain, clean, and analyze the data; check for accuracy and reasonableness (dirty data will degrade forecasts).

  • Step 4: Select a forecasting technique.

  • Step 5: Produce the forecast.

  • Step 6: Monitor forecast error and adjust as needed; forecasting is a cycle, not a single event, because errors reveal how the model can be improved.

  • Forecast error: the difference between the actual value and the forecast value.

    • If the error falls outside established boundaries (e.g., you forecast 100 bikes and sell 80), a corrective action may be required (revise the model or use a different technique).
    • Forecast accuracy is assessed using historical performance through accuracy metrics.

Forecast accuracy metrics (MAD, MSE, MAPE)

  • There are three common forecast accuracy metrics:

    • Mean Absolute Deviation (MAD):
      MAD = rac{

      \sum{t=1}^{n} \,|At - F_t|
      \n }{n}

    • At = actual value at time t; Ft = forecast at time t; n = number of periods.

    • Mean Squared Error (MSE):
      MSE = rac{\sum{t=1}^{n} (At - F_t)^2}{n - 1}

    • Squaring emphasizes larger errors.

    • Mean Absolute Percentage Error (MAPE):
      MAPE = rac{1}{n} \,\sum{t=1}^{n} \, \left| \frac{At - Ft}{At} \right| \,\times 100

    • Expresses error as a percentage of the actual value.

  • Worked example (illustrative data from transcript):

    • Data (periods 1–5):
    • Period 1: A1 = 110, F1 = 107, error e1 = A1 - F_1 = 3 (but transcript presents -3 for F1 from A1; absolute error = 3).
    • Period 2: A2, F2, e_2, etc. (not all values shown in transcript).
    • Provided summary values in the transcript:
    • Sum of absolute errors:
      \sum{t=1}^{5} |At - F_t| = 13
    • Sum of squared errors:
      \sum{t=1}^{5} (At - F_t)^2 = 9.75
    • Sum of absolute percentage errors (per-period, in percentage):
      \sum{t=1}^{5} \left| \frac{At - Ft}{At} \right| \,\times 100 \,\approx 11.23
    • From these sums:
    • MAD: MAD = rac{13}{5} = 2.6
    • MSE: first compute the mean of squared errors with n - 1 in the denominator: MSE = rac{9.75}{5 - 1} = rac{9.75}{4} \,=\ 2.4375
      • The transcript notes the sum of squared errors as 9.75, leading to an MSE of 2.4375 (rounded as appropriate).
    • MAPE: using the transcript value, MAPE = \frac{11.23}{5} \approx 2.246\% \approx 2.25\%.

  • Interpretation:

    • MAD provides average magnitude of errors in the original units.
    • MSE penalizes larger errors more strongly due to squaring.
    • MAPE expresses error as a percentage of actual demand, enabling comparison across items or time periods.
    • These metrics are used to compare different forecasting techniques using the same data; the goal is to minimize these error measures.

Forecasting approaches: qualitative vs quantitative

  • Qualitative forecasting:
    • Relies on soft information and human judgment (e.g., CEO intuition, expert opinion, market perception).
    • Difficult to quantify precisely and often used to supplement quantitative methods.
  • Quantitative forecasting:
    • Relies on hard data and statistical methods; increasingly dominant in the era of big data.
    • Provides objective, reproducible results.
  • Interaction of approaches:
    • Qualitative insights can inform the selection or adjustment of quantitative models (e.g., recognizing anomalies or likely future events).
    • Example from transcript: airline passengers may spike in October due to a World Series-related event in a given year; this kind of one-off effect may not repeat next year, so qualitative judgment helps avoid overfitting to such an anomaly.

Time series forecasting concepts (patterns and components)

  • Time series: a sequence of observations taken at regular time intervals (e.g., customers served per hour, sandwiches made per day).
  • Time series forecast aims to estimate future values by identifying patterns in recent observations.
  • Common time series behaviors/patterns (shown in transcript):
    • Trend: a general upward or downward movement over time (e.g., an upward trend in demand).
    • Irregular variation: a short-lived, unexplained spike (e.g., a weather event causing a sudden surge in demand).
    • Random variation: unpredictable fluctuations without a discernible pattern.
    • Seasonality: regular, predictable patterns that repeat within fixed periods (e.g., higher demand in December, or weekly patterns such as supermarket sales peaking on weekends).
    • Cycles: repeated rises and falls with a length that is not fixed to a specific period (e.g., cycles in housing markets or inflationary periods).
  • Examples discussed in transcript:
    • Half-hourly electricity demand in England and Wales showing seasonality/recurrence over weeks.
    • Monthly sales of new single-family houses in the US from 1973–1995, showing strong seasonality and cyclic behavior with periods roughly 6–10 years.
    • Supermarket sales showing weekly seasonality (higher spending on Fridays/Saturdays) and holiday-related seasonal increases.

Forecasting techniques introduced in the module

  • Naive forecasting:

    • Very simple: the forecast for the next period equals the last observed value.
    • Works best for stable time series with little to no trend or seasonality; easy to understand and implement.
    • Example (transcript): for a stable series with a last data point of 64 (period 5), the naive forecast for period 6 is 64.

  • Averaging forecasts (includes moving average variants, discussed next):

    • Averages recent observations to smooth fluctuations and generate forecasts.
    • Useful when the series tends to vary around a central level.

  • Moving average (simple moving average):

    • Forecast is the average of the most recent m actual observations.
    • Formula:
      \hat{y}t = \frac{1}{m} \sum{i=1}^{m} y_{t-i}
    • Example from transcript: using a 5-period moving average with data back to period t-5 yields a forecast of 61 (compared to naïve 64).
    • Characteristics:
    • As new data become available, older data are dropped; tends to lag the actual values.
    • Smoothing effect increases with larger m; decreases responsiveness to changes in the series.
    • Treats all data points equally (equal weighting).

  • Weighted moving average:

    • More recent observations get higher weights; weights sum to 1.
    • Rationale: recent data are typically better predictors than older data.
    • General formula:
      \hat{y}t = \sum{i=0}^{m-1} wi \, y{t-i}, \, \quad \sum{i=0}^{m-1} wi = 1
    • Example from transcript (illustrative weights): use the last three data points with weights 0.5 (most recent), 0.3, and 0.2; forecast becomes 0.5 yt + 0.3 y{t-1} + 0.2 y_{t-2} = 60.4 (rounded to 60 for practicality).

  • Exponential smoothing (also called single exponential smoothing):

    • A weighted moving average where weights decline exponentially for older observations; uses a smoothing constant α ∈ (0, 1).
    • Concept: forecast for the next period equals previous forecast plus a portion of the previous forecast error.
    • Formula (as described in transcript):
      F{t+1} = Ft + \alpha (At - Ft) = (1 - \alpha) Ft + \alpha At
    • Example from transcript (α = 0.4):
    • Period 2 forecast: using naive, F_2 = 60 (from period 1).
    • Period 3 forecast: F3 = F2 + 0.4 (A2 - F2) = 60 + 0.4 (65 - 60) = 62.
    • Period 4 forecast: F4 = F3 + 0.4 (A3 - F3) = 62 + 0.4 (55 - 62) = 59.2.
    • Period 5 forecast: F5 = F4 + 0.4 (A4 - F4) = 59.2 + 0.4 (58 - 59.2) = 58.72.
    • Period 6 forecast (as stated in transcript): F_6 = 60.83 (following the same rule with the given α and data).
    • Characteristics:
    • More responsive than simple moving averages for small α; smoother for small fluctuations when α is small.
    • The choice of α influences sensitivity to recent changes.

  • Overall takeaway from the techniques:

    • Four techniques were demonstrated: Naive, Moving Average, Weighted Moving Average, and Exponential Smoothing.
    • No single method is best for all data; model performance depends on the time series structure (trend, seasonality, variability).
    • The next module would cover more time series forecasting techniques and their applications.

Connections to practice and real-world relevance

  • Forecasts are used to balance supply and demand, plan capacity, manage inventory, and schedule production.
  • Accuracy metrics (MAD, MSE, MAPE) provide quantitative means to compare forecasting methods and choose an appropriate technique for a given dataset.
  • Time series components (trend, seasonality, cycles, irregular and random variations) guide which forecasting method to apply. For example:
    • Strong seasonality might favor methods that explicitly account for seasonal effects.
    • High random variation may require smoother methods with more history, or alternative approaches that can handle noise.
  • Ethical and practical implications:
    • Relying on forecasts without considering external changes (policy shifts, technology changes) can lead to poor decisions.
    • Overreliance on a single model can reduce resilience; combining qualitative insights with quantitative forecasts can improve robustness.
    • Simpler models (like the naive forecast) can be useful benchmarks and are often easier to explain to stakeholders, even if they are not the most accurate.

Quick recap: key takeaways

  • Forecasting aims to predict future values and to support decision-making under uncertainty.
  • Forecast quality depends on: usefulness of the forecast, accuracy (error magnitude), timeliness, reliability, meaningful units, and simplicity.
  • Forecasting involves a cycle: define purpose, determine horizon, collect and clean data, choose method, forecast, monitor and adjust.
  • Error metrics MAD, MSE, and MAPE help quantify forecast accuracy and compare methods.
  • Time series methods include naive forecasting, moving averages (simple and weighted), and exponential smoothing; each has trade-offs in simplicity, lag, and responsiveness.
  • Real-world data exhibit trend, seasonality, cycles, and irregular/random variations; appropriate modeling requires recognizing these patterns and choosing suitable techniques.

End of notes