demand forecasting

• Demand Forecasting: The process of estimating future customer demand for a

product or service.
• Independent Demand: Demand that is not directly dependent on the demand for
other products (e.g., a finished car).
• Dependent Demand: Demand that is driven by the demand for another product or
service (e.g., car tires needed to build a car).
• Qualitative Forecasting: Forecasting methods that rely on expert opinion, surveys, and
other non-numerical information.
• Quantitative Forecasting: Forecasting methods that use historical data and
mathematical models to predict future demand.
• Causal Methods: Forecasting techniques that attempt to identify cause-and-effect
relationships between demand and other factors.
• Time Series Methods: Forecasting techniques that use historical demand data over
time to predict future demand.
• Time Horizon: The length of time into the future that a forecast covers (short-term,
medium-term, long-term).
• Aggregate Forecast: A forecast for a group of products or customers.
• Disaggregate Forecast: A forecast for a specific product or customer.
• Systematic Component: Underlying patterns in the time series (trend, level,
seasonality).
• Random Component: Fluctuations in the data that cannot be explained by underlying
patterns (noise).
• Level (Lt): The current height of the time series.
• Trend (Tt): The long-term change over time (slope) of the time series.
• Seasonal Effect (St): A periodic repetition of a particular pattern in the time series.
• Error Measures: Metrics used to quantify the accuracy of a forecast (MSE, MAD, MAPE).
• Additive Configuration: A time series model where components are added together:
Dt = (Lt + Tt) + St + Rt
• Multiplicative Configuration: A time series model where components are multiplied:
Dt = (Lt + Tt)StRt
• Rolling Time Horizon: Regularly updating sales forecasts to ensure their continued
relevance and reliability.
Principles of Demand Forecasting

• Motivation: Demand forecasting is crucial because management inherently deals with

the future.
• Dependent vs. Independent Demand: It's important to distinguish between these.
• Independent Demand: Direct orders from end customers. This needs to be predicted.
• Dependent Demand: Driven by the demand for other products (e.g., demand for car
tires depends on the demand for cars). This can be computed using material
requirements planning (MRP).
• Demand vs. Sales: Demand is not the same as sales. "Not all demand will result in
sales," and "Not all sales comes from demand."
• External Factors: Many external factors (market trends, economic indicators,
demographics) influence demand.
2. Quality of the Forecast & Error Measures
• Judging Forecast Quality: The quality of a forecast can only be accurately assessed
after the actual demand is realized.
• Historical Data: A core assumption is that models that perform best at predicting the
past (using historical data) will be most suitable for predicting future demand.
• Forecast Error: The forecast error (Et) is the difference between the forecast (Ft) and
the actual demand (Dt): Et = Ft - Dt.
• Key Error Measures:Mean Squared Error (MSE): MSE = (1/T) * Σ(Et)^2. Penalizes large
errors more severely. Useful when the cost of a large error is high.
• "Especially useful in situations where the cost of a large error outweighs the gains from
having a highly accurate forecast."
• Mean Absolute Deviation (MAD): MAD = (1/T) * Σ|Et|. Useful when the cost of an
error is directly proportional to the magnitude of the error.
• "Especially useful in situations where the coast of a forecast error is directly
proportional to the magnitude of the error."
• Mean Absolute Percent Error (MAPE): MAPE = (1/T) Σ(|Et| / Dt) 100. Expresses error as
a percentage of actual demand. Similar to MAD, but relative to the data's magnitude.
3. Qualitative Forecasting Methods
• Relies on expert opinions and judgments. Methods include:
• Expert View: Seeking advice from established experts regarding their opinions and
expectations.
• Delphi Method: Structured, multi-expert survey that brings together specialists from
different companies and institutions to achieve consensus.
4. Quantitative Forecasting Methods
• Causal Methods: Assume a relationship between the demand (dependent variable, Y)
and one or more independent/explanatory variables (X).
• Linear Regression: A common causal method assuming a linear relationship: ŷ = a + bx.
• Optimal values for 'a' (y-intercept) and 'b' (slope) are calculated to minimize the MSE.
• b = (Σ(xiyi) - n*x̄*ȳ) / (Σ(xi^2) - n*x̄^2)
• a = ȳ - b*x̄
• Important Notes on Linear Regression:
1. Verify the data visually to see if a linear relationship is plausible. If not, curvilinear
analysis may be necessary.
2. When predicting y-values, it is important not to go far beyond the range for which we
actually have historical data.
3. Deviations around the least-squares line are assumed to follow a normal distribution
• Time Series Methods: Analyze historical data indexed by time, assuming past patterns
will repeat in the future. Suitable for short- and medium-term predictions.
• Components of a Time Series:Systematic Component: Underlying patterns to be
understood.
• Random Component: Fluctuations that cannot be explained by underlying patterns
(noise).
• "A time series can always be decomposed in two main components: a systematic and
a random component"
• Time Series Decomposition: Break down the time series into:
• Level (Lt): Current height of the time series.
• Trend (Tt): Long-term change over time (slope).
• Seasonal Effect (St): Periodic repetition of a pattern.
• Residual Term (Rt): Random noise, cannot be forecasted.
• Additive vs. Multiplicative Configurations:Additive: Dt = (Lt + Tt) + St + Rt
• Multiplicative: Dt = (Lt + Tt) St Rt
• Linear Regression for Time Series: Applying linear regression with time as the
independent variable.
• Moving Average: Forecast is the average of the 'r' most recent demand entries.
• FT+1 = (DT + DT-1 + ... + DT-(r-1)) / r
• Only forecasts one period into the future. Value of 'r' must be chosen carefully.
• Simple Exponential Smoothing (Brown Method): Weighted average of past forecast
and current demand.
• Ft+1 = Ft + α(Dt - Ft) = αDt + (1 - α)Ft
• 'α' is the smoothing constant (0 to 1). Higher α = more responsive to recent data.
• Suitable when demand doesn't exhibit a trend or seasonality.
• Trend-Corrected Exponential Smoothing (Holt's Method): Accounts for linear trends.
Decomposes time series into level (Lt) and trend (Tt).
• Ft+1 = Lt + Tt
• Lt = αDt + (1 - α)(Lt-1 + Tt-1)
• Tt = β(Lt - Lt-1) + (1 - β)Tt-1
• Uses two smoothing constants: α (level) and β (trend).
• Trend- and Seasonality-Corrected Exponential Smoothing (Winters' Method):
Handles both trends and seasonality.
• Decomposes time series into level (Lt), trend (Tt), and seasonal factors (St).
• Uses three smoothing constants: α (level), β (trend), and γ (seasonality).
• Distinguishes between additive and multiplicative models.
• Requires initializing level (L0), trend (T0), and seasonal factors (S1...Sp).
5. Model Selection and Implementation
• Overview of Time Series Methods (Decision Tree):No Trend, No Seasonality: Moving
Average, Simple Exponential Smoothing
• Trend, No Seasonality: Holt's Method, Linear Regression
• Trend and Seasonality: Winters' Method
6. Feedback and Reflection Decisions should not solely be based on the outcome of
quantitative models. The assumption that patterns from the past will continue in the future
should constantly be questioned. In practice, often a forecast is made based upon
consensus between various models. The result of the forecast exercise is the input for the
decision making process (and not the decision itself).
Key Takeaways:
• Demand forecasting is essential for effective supply chain management.
• Different forecasting methods exist, each with strengths and weaknesses.
• Model selection depends on the characteristics of the demand data (trend,
seasonality).
• Forecasts are inherently inaccurate, so it's vital to measure and understand forecast
error.
• Quantitative methods should be complemented by qualitative insights and expert
judgment.
• Collaboration among stakeholders improves forecast accuracy.Glossary of Key Terms

• Additive Configuration: A time series model where the components (level, trend,
seasonal effect, and residual) are added together to create the final demand value.
Formula: Dt = (Lt + Tt) + St + Rt
• Aggregate Forecast: A demand forecast that combines data for multiple products,
customers, or geographic areas. It provides a broad overview of demand patterns.
• Causal Methods: Quantitative forecasting techniques that attempt to identify
cause-and-effect relationships between demand and one or more independent
variables. Regression analysis is a common example.
• Delphi Method: A qualitative forecasting technique that involves a structured survey
process with multiple experts to reach a consensus forecast.
• Dependent Demand: Demand for a product or component that is directly derived
from the demand for another product (e.g., demand for tires depends on the demand
for cars).
• Disaggregate Forecast: A demand forecast that is specific to a particular product,
customer, or geographic area. It provides detailed demand insights.
• Demand Forecasting: The process of estimating future customer demand for products
or services.
• Error Measures: Metrics used to quantify the accuracy of a forecasting model by
comparing forecasted values to actual demand values. Examples include MSE, MAD,
and MAPE.
• Holt's Method: (Trend-Corrected Exponential Smoothing): An exponential smoothing
technique that accounts for both level and trend in a time series, suitable for data with
a linear trend but no seasonality.
• Independent Demand: Demand for a finished product that is not directly dependent
on the demand for any other product.
• Level (Lt): In time series decomposition, the average value of the series at a particular
point in time, representing the baseline demand.
• Linear Regression: A causal forecasting method that uses a linear equation to model
the relationship between a dependent variable (demand) and one or more
independent variables.
• MAD (Mean Absolute Deviation): A measure of forecast error that calculates the
average absolute difference between forecasted and actual values.
• MAPE (Mean Absolute Percent Error): A measure of forecast error that calculates the
average percentage difference between forecasted and actual values, relative to the
actual values.
• Moving Average: A time series forecasting technique that calculates the forecast for
the next period as the average of the r most recent demand entries.
• MSE (Mean Squared Error): A measure of forecast error that calculates the average of
the squared differences between forecasted and actual values.
• Multiplicative Configuration: A time series model where the components (level, trend,
seasonal effect, and residual) are multiplied together to create the final demand value.
Formula: Dt = (Lt + Tt)StRt
• Qualitative Forecasting: Forecasting methods that rely on expert opinions, market
research, surveys, and other non-numerical data.
• Quantitative Forecasting: Forecasting methods that use historical data and
mathematical models to predict future demand.
• Random Component: The unpredictable, irregular fluctuations in a time series that
cannot be explained by the systematic components. Also known as "noise" or "residual
variation."
• Rolling Time Horizon: A forecasting approach where forecasts are continuously
updated as new data becomes available, extending the forecast horizon forward in
time.
• Seasonal Effect (St): A recurring pattern in a time series that repeats at regular intervals
(e.g., monthly, quarterly, or yearly).
• Simple Exponential Smoothing (Brown Method): A time series forecasting technique
that uses a weighted average of past observations, with more recent observations
receiving higher weights. Suitable for data with no trend or seasonality.
• Systematic Component: The predictable, underlying patterns in a time series, including
trend, level, and seasonality.
• Time Horizon: The length of time into the future that a forecast covers (e.g.,
short-term, medium-term, long-term).
• Time Series Methods: Quantitative forecasting techniques that analyze historical data
patterns over time to predict future demand.
• Trend (Tt): The long-term direction or movement of a time series, representing the
general increase or decrease in demand over time.
• Winters' Method: (Trend- and Seasonality-Corrected Exponential Smoothing): An
exponential smoothing technique that accounts for both trend and seasonality in a
time series, suitable for data with both linear trends and seasonal patterns.