MBA/MBA Semester 1/Statistics for Management/Module 5: Forecasting Techniques

Module 5: Forecasting Techniques — Notes

Lesson 1: Learning Objectives

  • Correlation and regression analysis; measuring relationships with covariance and Pearson R.
  • Practical uses of covariance and correlation: varieties, types of coefficient and types of correlation.
  • Regression modelling essentials and types; time series analysis and its types.

Lesson 2: 5.1 Correlation and Regression Analysis (SLM p.140)

  • Correlation analysis determines whether two variables are related and how strongly. Regression analysis finds the degree and nature of the relationship between a dependent variable Y and one or more independent variables.
  • The most popular form is linear regression (simple / ordinary least squares), which fits a line of best fit; nonlinear models are more intricate.

Lesson 3: 5.1.1 Covariance and Correlation Overview (SLM p.140-141)

  • Data often come as pairs (price and demand, money supply and inflation).
  • Croxton and Cowden: correlation is the tool for discovering and measuring a quantitative relationship; A.M. Tuttle called it the "analysis of covariation".
  • Correlation measures the degree of linear relationship; it makes no dependent / independent distinction and can give ludicrous findings if the underlying relationship is ignored.

Lesson 4: 5.1.2 Measuring Relationships with Covariance (SLM p.141)

  • Correlation gauges the degree of a relationship; covariance gives its direction. Both are positive when variables move together and negative when they move in opposite directions.
  • The correlation coefficient lies between -1 and +1.
  • Covariance = Σ[(x - x̄)(y - ȳ)] ÷ (n - 1). Covariance depends on scale, so it is limited when comparing datasets.

Lesson 5: 5.1.3 Pearson R: Relationship Analysis (SLM p.142)

  • Pearson's correlation coefficient assesses the direction and degree of a relationship between two continuous variables, using covariance; the most effective approach for evaluating relationships.
  • Important premises (markers): each instance stands alone from the others (independence); a scatterplot confirms a linear connection (a straight line in the plot); a roughly rectangular shape should appear in the residuals scatter plot.
  • Features (markers): 0 = no correlation, range +1 (perfect positive) to -1 (perfect negative); independent of units of measurement, so comparable across scales; symmetric - the correlation of X with Y equals that of Y with X.
  • Correlation degrees: Perfect (values close to ±1: a rise or fall in one is mirrored in the other); High (±0.50 to ±1); Moderate (±0.30 to ±0.49); Low (below 0.29); No correlation (0).

Lesson 6: 5.1.4 Practical Uses of Covariance and Correlation (SLM p.142-143)

  • Correlation studies how much two or more variables fluctuate in relation to one another. Caution example: India's population and average global temperature both rose over 50 years but are manifestly unrelated.
  • Coefficient lies between -1 and +1. Positive: variables move in the same direction (e.g. crop yield and fertiliser applied). Negative: opposite directions (e.g. a commodity's price and its demand). About zero: no correlation. Near ±1: strongly linear relationship.
  • How managers benefit (markers): (1) to determine how different factors and decision variables relate to one another; (2) if two variables are correlated, to estimate the value of one given the other; (3) to comprehend market dynamics and economic conduct; (4) to greatly minimise uncertainty in decision-making.
  • Correlation analysis helps decisions by estimating the effect of decision variables (price, production process, advertising and promotion) on objectives (costs, sales, market share, customer satisfaction, competitive price). It removes subjectivity, but it only shows that variables move together - not necessarily a cause-and-effect relationship; logical understanding of the business environment is needed.
  • Finance uses: correlation supports asset diversification; covariance is used in portfolio risk analysis.

Lesson 7: 5.1.5 Correlation Varieties (SLM p.143-145)

  • Correlation can be partial or total, simple or multiple, linear or nonlinear, positive or negative. Studied by plotting x-y graphs (line or scatter diagrams) or by computing the coefficient algebraically. Coefficients: coefficient of determination, Spearman's rank correlation and Karl Pearson's correlation coefficient.
  • Positive or negative: positive (direct) = both rise or fall together; negative (inverse) = one rises as the other falls. A scatter diagram with all points on a straight line indicates perfect correlation.
  • Simple or multiple: simple = two variables with little outside effect (e.g. price changes vs sales of a price-sensitive product in a stable market); multiple = several variables affect one another, so each pair must be investigated.
  • Partial or total: partial = how two variables vary while controlling for the influence of other factors (e.g. wheat yield vs fertiliser ignoring the effect of pesticides); total = all relevant variables allowed to vary together (impractical as variables increase).
  • Linear and nonlinear: linear = a constant ratio of change between the variables (straight line on the graph); nonlinear = the rate of variation changes. Most coefficients assume linearity, so plot a scatter diagram first.

Lesson 8: 5.1.6 Karl Pearson's Coefficient Introduction (SLM p.145-148)

  • For grouped data a bivariate frequency distribution (rows X, columns Y; cell (i, j) = joint frequency) is used, with class marks mX and mY, marginal frequencies fX and fY and joint frequencies fXY. Worked example uses assumed mean 1250 and scaling factor 500 (step-deviation method).
  • Definition: the correlation coefficient r (product moment correlation coefficient) measures the degree of connection between X and Y: r = Cov(X, Y) / (σX σY), where Cov(X, Y) = (1/n) Σ(X - X̄)(Y - Ȳ) and n is the number of pairs. r is a dimensionless number between +1 and -1. Shortcut (assumed mean) procedure uses deviations U and V, e.g. advertising expenditure (X) vs revenue (Y) over ten years.
  • Interpretation: r = +1 perfect positive (all points on a rising line), r = -1 perfect negative, r = 0 no linear correlation. (Xi - X̄)(Yi - Ȳ) is positive when both are on the same side of their means, so r is positive when X and Y tend to be above or below their means together and negative when they fall on opposite sides.
  • r is unaffected by measurement units (and by change of origin and scale).

Lesson 9: 5.1.7 Spearman Rank Correlation Primer (SLM p.148-151)

  • Data are often given as rankings (e.g. experts ranking candidates on sincerity, loyalty, integrity, tact, initiative, or dance contestants) because some qualities cannot be measured in numbers. The rank correlation coefficient (Spearman's rS) measures how much two sets of ranks agree.
  • Ranks given: example of ten students' ranks in mathematics (X) and physics (Y); rS computed from rank differences d (rS = 1 - 6Σd² / [n(n² - 1)]); a high value means performance in the two subjects is highly correlated.
  • Ranks not given: assign ranks yourself (biggest item gets rank 1, next biggest rank 2, and so on) in X (R1) and Y (R2), then compute rS.
  • Spearman is used when data are ordinal or nonlinear (rank-based).

Lesson 10: 5.1.8 Pearson and Spearman Correlation Uses (SLM p.151-153)

  • Pearson (linear correlation) measures the linear link between two continuous variables. Check first: random sampling; continuous quantitative variables; paired data; independent observations; normal distribution; linear relationship; no outliers. r near +1 = linear positive dependence; near 0 = no linear relationship; near -1 = linear negative dependence. Example use: country-specific meat intake vs life expectancy.
  • Spearman is a rank-based measure of a monotonic relationship; ranking example [58, 70, 40] becomes [2, 1, 3]. Frequently used for data containing outliers. Assumptions: random sampling, paired and independent observations, monotonic relationship; variables may be continuous or ordinal.

Lesson 11: 5.1.9 Regression Modelling Essentials (SLM p.153-154)

  • A statistical model captures the systematic behaviour of the data; what is left is random error (the unpredictable part plus what the model could not explain). If errors are assumed Normally distributed we can state confidence levels and intervals.
  • If a scatter diagram clusters around a curve, the variables are correlated; a straight line is linear regression, otherwise curvilinear regression. "Best fit" is found by Legendre's principle of least sum of squares of deviations (minimum squared error criterion). Measuring deviations in the Y direction gives the regression of Y on X; in the X direction, the regression of X on Y - two different best-fit lines.
  • Morris Myers Blair: regression is the average relationship between two or more variables expressed in the data's original units.
  • Regression is widely used, but it only gives mathematical measures from the available data: it does not verify cause and effect or say which variable depends on which. It helps by: giving mathematical relationships between variables; providing a model for cause-and-effect relationships; analysis, forecasting and estimation; establishing theories (supply and demand, money supply and expenditure, inflation and interest rates, sales and promotion cost, productivity and profitability, employee health and absenteeism).

Lesson 12: 5.1.10 Least Square Method Overview (SLM p.154-156)

  • The least-square method finds the "best" straight line: it minimises the sum of squared residuals. It determines constants a and b of Y = a + bX; r can then be computed.
  • Normal equations for Y = a + bX: ΣY = na + bΣX and ΣXY = aΣX + bΣX². Solving them gives a and b; the fitted line is the regression of Y on X. If X is dependent and Y independent we fit X = a + bY instead.
  • Regression analysis builds a mathematical model of the relationship between variables to predict the dependent variable from the independent one.
  • Parabolic trend: Yt = a + bt + ct², with three normal equations for a, b, c. Choose an origin year so that ΣX = 0 (X = t - origin); then ΣX³ = 0 too and the equations simplify.

Lesson 13: 5.1.11 Curve Fitting Basics (SLM p.156-157)

  • Repeats the regression modelling ideas: a model captures the systematic part and leaves only random errors; points clustering around a curve imply correlation; straight line = linear regression, otherwise curvilinear; best fit by Legendre's least sum of squares (minimum squared error criterion); deviations in Y direction give regression of Y on X.
  • Curve fitting finds a smooth function approximating the data; polynomial fitting is used for non-linear trends.

Lesson 14: 5.1.12 Model Evaluation Techniques (SLM p.157)

  • Model evaluation metrics determine how well a model fits the data, compare models, aid model selection and forecast the accuracy of predictions.
  • R² (coefficient of determination) shows the fraction of variance explained; a good model has R² close to 1.

Lesson 15: 5.1.13 Standard Error of Estimate (SLM p.157-158)

  • The standard error of estimate (SE of Y) measures the variance (dispersion) of observed Y values around the regression line, i.e. residual variability. A lower SE indicates a stronger fit.
  • Regression coefficient of X on Y (bxy): change in X for a unit change in Y. Regression coefficient of Y on X (byx): change in Y for a unit change in X. Each has a formula using deviations from actual means and another using deviations from assumed means. The regression coefficient is also called the slope coefficient.

Lesson 16: 5.1.14 Coefficient of Determination (SLM p.159)

  • The percentage of variation in the dependent variable that can be predicted from the independent variable(s); symbol R² or r² ("R squared"). Used in models that forecast future events or test hypotheses; shows how well observed results are reproduced by the model as a share of total variation explained.
  • With simple linear regression and an intercept, r² is the square of the sample correlation r; with more regressors R² is the square of the multiple correlation coefficient. It usually lies between 0 and 1; it can be negative when predictions were not produced by a fitting process on those data, or for non-linear fits or regression without an intercept (the data mean then fits better than the fitted function).
  • R² is more intuitive than MAE, MAPE, MSE and RMSE because it is expressed as a percentage (the others have arbitrary ranges), and it is more robust against bad fits than SMAPE.

Lesson 17: 5.2 Time Series Analysis (SLM p.159-160)

  • Time series analysis is a methodical way to find and separate time-related patterns. The four most prevalent patterns are cyclic, trend, seasonal and horizontal, with a random component on top; the breakdown is used for forecasting.
  • A more accurate approach is to identify patterns with autocorrelation - correlation between values of the same variable at different time lags. Completely random data: autocorrelations near zero, fluctuating both sides. Autocorrelation that drops slowly to zero with more than two or three significantly non-zero values indicates a trend, which can be removed by differencing consecutive values (numerical differentiation).
  • Definition: a time series is data from observing a response variable at periodic points in time; Yi denotes the value at time i.
  • Two primary goals (markers): (1) to examine the data's historical behaviour; (2) to create future projections - history shows the effects of different pressures, aiding forecasting and planning.

Lesson 18: 5.2.1 Time Series Analysis Introduction (SLM p.160-161)

  • Time series analysis reveals the fundamental reasons behind long-term trends or systemic patterns; visualisations help spot seasonal trends and their causes.
  • Time series forecasting estimates the probability of future events from data analysed at regular intervals; it is part of predictive analytics and can highlight expected changes such as seasonality or cyclic behaviour.
  • Example: Des Moines Public Schools analysed five years of student achievement data to identify at-risk students and monitor progress over time.
  • (The lesson page also repeats the coefficient of determination text from Lesson 16.)
  • A secular trend (or simply trend) is the general tendency of data to rise, fall or stay constant over an extended period. Most business and economic series show a tendency to increase or decline over years: population, bank deposits, industrial and agricultural production, deficit finance tend to increase.
  • A series may also show a declining trend: demand for a commodity being replaced by another (cotton clothing, coarse grains such as jowar and bajra) or the death rate falling with better medical facilities.
  • Trend changes are linked to basic drivers: population growth, technological advancement, shifts in the makeup of the industrial process.

Lesson 20: 5.2.3 Seasonal Variation Analysis (SLM p.162)

  • Seasonal variations are cycles on short timescales (usually under a year: weekly, daily, monthly). A series shows both seasonal and cyclical fluctuations if the interval between observations is one year or less; if observations are more than a year apart there are no seasonal fluctuations.
  • Main causes: (1) Climatic conditions - e.g. woollen clothing sales peak in November-December; timely rain boosts farm output and agricultural prices are lowest at harvest; (2) Customs and traditions - e.g. clothes and ornament sales peak in the marriage season, sweets sell at Diwali.
  • Seasonal effect is analysed with the ratio-to-trend method (among others).

Lesson 21: 5.2.4 Cyclical and Irregular Variations (SLM p.162-163)

  • Cyclical component (markers): (1) most economic and commercial time series show cyclical changes, hence called trade or business cycles; (2) every trade cycle has four periods in order - boom, recession, depression, recovery; (3) the stages alternate in a regular manner, following one another in the prescribed order; (4) the period of cyclical variation is the span between two consecutive phases; (5) the duration is always more than a year, typically three to ten years.
  • Objectives of measuring cyclical variations: examine historical cyclical behaviour; forecast the impact of cyclical fluctuations to guide future business strategy.
  • Irregular variations are unsystematic and random.

Lesson 22: 5.2.5 Numerical Trend Analysis (SLM p.163-164)

  • Example 1: determine the trend of given data by the free-hand method (plot the data and draw a smooth line by eye).
  • Example 2: find trend values using three-yearly moving averages and show the trend line on a graph (compute each 3-year total, divide by 3 and place it against the middle year).
  • Numerical trend methods include curve fitting, moving averages and percentage change (all of the above). A moving average helps remove seasonality, forecast short-term trends and smooth out noise.

Lesson 23: 5.2.6 Measuring Cyclical and Irregular Variations (SLM p.165)

  • Cyclic variation: the data rise and fall over a period that is not fixed. The most popular technique is the residual approach: estimate the trend (T) and seasonal (S) components first, then remove them; irregular components are removed by moving averages (smoothing).
  • Steps: (1) get initial estimates of trend T and seasonal values S; (a) divide the series values Y by T and S to obtain the cyclic (C) and random (R) components (Y / (T x S) = C x R); (b) remove the random component using a 3- or 5-month moving average to get the cyclic component.
  • Other approaches: cyclic analysis method references; the straightforward percentage fluctuation method; fitting with harmonic analysis / sine function.
  • Irregular component (error term): the final component, a result of natural forces that no series can fully eliminate; no published technique quantifies it, but a little can be removed by averaging the indices. Multiplicative model: divide the other components out; additive model: subtract them.

Lesson 24: 5.2.7 Numerical Trend Analysis Application (SLM p.166-167)

  • Trend analysis examples help managers and stakeholders (including investors) use company information, e.g. public web data such as website traffic.
  • Example 1 (retailer): firm A, an online gift retailer, shows six months of internet traffic with an upward trend peaking on 20 December (holiday season), then roughly horizontal after the first week of January; a rival can benchmark against it and use it to forecast outcomes.
  • Example 2 (share prices): one-year comparison of X and Y. X trended upward (apart from a slight reversal in February), suggesting a worthwhile long-term portfolio addition; Y was horizontal for six months then declined (a loss if added). With a horizontal trend investors are cautious because the direction of the next move is hard to predict.
  • Other data (the business itself, the market, the economy) should be considered too; trend analysis is one of several tools for judging an asset's profitability.
  • Retailers use trend analysis to optimise pricing and inventory planning; it identifies long-term consumer behaviour.

Lesson 25: 5.2.8 Case Study - Retail Chain Sales Trend Analysis (SLM p.168-169)

  • Context: a retail chain with stores in several countries must examine the previous five years of sales data to find trends that guide inventory management, marketing initiatives and store expansion. Goal: numerical trend analysis to find patterns, seasonality and long-term trends by product category and location.
  • Actions (8 steps): (1) Gathering data - monthly sales by product type for five years, plus demographics, average transaction value, units sold, revenue. (2) Preparing data - clean errors, duplicates and missing values; organise in a database or spreadsheet. (3) Analysis of the time series - mean, median, variance, standard deviation per period. (4) Identifying trends - moving averages and linear regression; plot to see seasonality, cyclical patterns, upward or downward tendencies. (5) Analysis of numerical trends - annual growth rate (AGR) for year-on-year % change and compound annual growth rate (CAGR) for the average yearly growth. (6) Comparing and dividing - segment by demographics, store location and category; compare performance to find opportunities and weaknesses. (7) Predicting - trend analysis plus forecasting methods (machine learning, ARIMA, exponential smoothing). (8) Conclusions and suggestions - summarise patterns and seasonality; recommend pricing, marketing, inventory and store-operation improvements; share with store managers, marketing and senior management.
  • Conclusion: numerical trend analysis gives insight into sales performance and supports well-informed decisions for growth and profitability.

Lesson 26: Summary

  • Data often occur as pairs of figures (commodity price and demand, money supply and inflation, industrial growth and GDP, advertising expenditure and market share).
  • Graphs are usually line or scatter diagrams; correlation coefficients include the coefficient of determination, Spearman's rank and Karl Pearson's.
  • r lies between -1 and +1: +1 means all points on a rising line, -1 all points on a falling line, 0 no linear correlation.
  • Best fit by Legendre's principle of least sum of squares (minimum squared error criterion).
  • The standard error of estimate (SE of Y) measures the variability of observed Y values around the regression line.

Lesson 27: Glossary

  • Negative correlation: movement of the variables in opposite directions. Linear correlation: the change in one variable keeps a constant ratio to the change in the other. Regression: a basic technique for measuring or estimating the relationship among economic variables. Time series: data obtained by observing a response variable at periodic points in time. Standard error of estimate: the measure of variation around the computed regression line.

Lesson 28: Check Your Understanding

  • 20-question quiz (pass mark 0%); scored 100%. Full set in Quiz Bank Q45-64.

Lesson 29: Exercise (notebook)

  1. Discuss covariance and correlation in measuring relationships between variables. 2. Types of graphs (line / scatter) and the different correlation coefficients (determination, Spearman's rank, Karl Pearson's). 3. Application of Spearman's rank correlation when ranks are not given. 4. Compare and contrast Spearman's and Pearson's coefficients. 5. The statistical model in regression analysis and its role in determining the relationship between independent and dependent variables.

Lesson 30: Learning Activities

  1. Workshop on Pearson's coefficient: groups calculate r for a two-variable dataset (height and weight, study hours and GPA). 2. Groups calculate Karl Pearson's r for datasets (height and weight, temperature and ice cream sales), then discuss and interpret the strength and direction of the relationship.

Lesson 31: Further Readings and Bibliography

  • Levin, Rubin, Rastogi, Siddiqui, Statistics for Management (Pearson, 7th ed., 2016); Prem S. Mann, Introductory Statistics (Wiley India, 7th ed.); James, Witten, Hastie, Tibshirani, An Introduction to Statistical Learning with Applications in R (Springer, 2016); Srivastava et al., Quantitative Techniques for Managerial Decision Making; Levin and Kirkpatrick, Quantitative Approaches to Management; Budnik, McLeavey, Mojena, Principles of Operations Research; J.K. Sharma, Operations Research - Theory and Applications.

✨ Quick Revision Pack (Module 5)

60-second summary

  • Correlation measures the degree and direction of a linear relationship, range -1 to +1; unit-free and symmetric. Covariance shows only direction (scale dependent). Correlation ≠ causation.
  • Pearson r = Cov(X,Y)/(σXσY): continuous, linear data. Spearman rS = 1 - 6Σd² / [n(n² - 1)]: ranked/ordinal data, monotonic, outlier-resistant.
  • Regression: least squares minimises Σ(residuals)². Y = a + bX; normal equations ΣY = na + bΣX, ΣXY = aΣX + bΣX². byx and bxy are two different regression coefficients; r² = byx x bxy.
  • R² (coefficient of determination) = % of variation in Y explained by X; standard error of estimate = scatter of points around the line (lower = better).
  • Time series = Trend (T) + Seasonal (S) + Cyclical (C) + Irregular (R). Trend by free-hand, moving averages, least squares. Cyclical measured by the residual method (Y / (T x S) = C x R).

Diagrams

Components of a time series

flowchart TB
  Y["Time series Y"] --> T["Trend (long-term direction)"]
  Y --> S["Seasonal (within a year, repeating)"]
  Y --> C["Cyclical (business cycle, not fixed length)"]
  Y --> R["Irregular (random shocks)"]

Reading the correlation coefficient

flowchart LR
  A["-1
perfect negative"] --- B["-0.5
moderate negative"] --- C["0
no linear link"] --- D["+0.5
moderate positive"] --- E["+1
perfect positive"]

From scatter plot to forecast

flowchart TB
  A["Scatter plot"] --> B["Compute r"] --> C["Fit line by least squares"] --> D["Check R² and standard error"] --> E["Forecast Y for a given X"]

Memory tricks

💡 Tip

Seasonal vs cyclical: Seasonal = Short and Set (within a year, fixed calendar). Cyclical = Cycle of the economy (years, not fixed).

💡 Tip

r² = byx x bxy and r = ±√(byx x bxy) (the sign of r equals the sign of both regression coefficients).

❗ Important

R² = r² in simple linear regression. R² = 0.81 means 81% of the variation in Y is explained by X (r = ±0.9).

💡 Tip

Correlation degrees: |r| ≥ 0.50 high; 0.30-0.49 moderate; below 0.30 low; 0 none.

⚠️ Warning

A high correlation between two trending series (e.g. population and global temperature) can be spurious. Correlation cannot prove cause and effect.

📘 Note

Moving average: 3-year MA places each average against the middle year; it smooths noise and removes seasonality. A 3-year MA loses the first and last observation.

Formula sheet

Concept Formula
Covariance Σ(x - x̄)(y - ȳ)/(n - 1)
Pearson r Σ(x - x̄)(y - ȳ) / √[Σ(x - x̄)² Σ(y - ȳ)²]
Spearman rS = 1 - 6Σd² / [n(n² - 1)]
Regression line (Y on X) Y = a + bX, b = (nΣXY - ΣXΣY)/(nΣX² - (ΣX)²), a = ȳ - b x̄
byx, bxy byx = r σy/σx ; bxy = r σx/σy
Coefficient of determination R² = r² (simple regression)
Moving average (3-year) (Yt-1 + Yt + Yt+1)/3
Multiplicative model Y = T x S x C x R

Practice problems (with answers)

  1. Spearman: ranks of 5 students: X = 1,2,3,4,5; Y = 2,1,4,3,5. → d = -1,1,-1,1,0; Σd² = 4; rS = 1 - 6(4)/[5(24)] = 1 - 24/120 = 0.80.
  2. Regression: ΣX = 15, ΣY = 40, ΣXY = 135, ΣX² = 55, n = 5. → b = (5x135 - 15x40)/(5x55 - 225) = 75/50 = 1.5; a = 8 - 1.5x3 = 3.5; Y = 3.5 + 1.5X. Forecast at X = 6: 12.5.
  3. r from regression coefficients: byx = 0.8, bxy = 0.45 → r = √(0.8 x 0.45) = √0.36 = 0.6 (positive, both coefficients positive).
  4. R²: r = -0.7 → R² = 0.49 (49% of the variation explained).
  5. 3-year moving average: sales 40, 46, 52, 50, 58 → MA = (40+46+52)/3 = 46 (for year 2), (46+52+50)/3 = 49.3 (year 3), (52+50+58)/3 = 53.3 (year 4).
  6. Seasonal index idea: if January sales are 20% above the monthly trend, the seasonal index = 120.
  7. Short answer (5 marks): Difference between correlation and regression. → Correlation measures strength/direction, no dependent variable, symmetric; regression gives an equation to predict one variable from another and treats X and Y asymmetrically.
  8. Classify: woollen sales peak in Nov-Dec → seasonal. Slump after a boom lasting years → cyclical. A sudden strike → irregular. Steady rise of bank deposits over decades → trend.