Time Series Basics - Business Analytics

SECTION 1

Historical Context & Motivation

The systematic analysis of data ordered by time is one of the oldest quantitative pursuits in human history. Long before the term time series analysis was coined, merchants, astronomers, and government officials recorded sequences of observations—crop yields, tidal patterns, commodity prices—hoping to detect regularities that would improve future decisions. In the business world, the ability to anticipate demand, revenue fluctuations, and economic cycles has always conferred a competitive advantage. The formalization of time series methods transformed this intuitive practice into a rigorous discipline, providing managers and analysts with a shared mathematical language for extracting trend, seasonality, and cyclical behavior from temporal data.

1927 • Yule's Autoregressive Model

George Udny Yule proposed the autoregressive (AR) model to explain sunspot numbers, demonstrating that a value in a series could be modeled as a linear function of its own past values.

1937 • Slutzky's Moving-Average Insight

Eugen Slutzky showed that summing random shocks through a moving-average (MA) process could generate realistic-looking economic cycles, laying a foundation for stochastic time series theory.

1970 • Box–Jenkins ARIMA Framework

George Box and Gwilym Jenkins published their landmark text unifying AR and MA models into the ARIMA framework, providing a systematic identify-estimate-diagnose methodology widely adopted in business forecasting.

1982 • ARCH/GARCH for Volatility

Robert Engle introduced the ARCH model to capture time-varying volatility in financial returns, a breakthrough that earned him the Nobel Prize in Economics in 2003.

2017 • Deep Learning & Prophet

Facebook released Prophet, an open-source forecasting tool, while recurrent neural networks (LSTMs) and later Transformers began outperforming classical models on complex, high-dimensional time series tasks in industry.

Throughout this evolution, the central question has remained the same: given a sequence of observations collected over time, how can we decompose and model the underlying structure to generate reliable forecasts? This lesson introduces the foundational vocabulary, decomposition techniques, and mathematical building blocks that every business analyst needs before diving into more advanced predictive models.

SECTION 2

Core Principles & Definitions

A time series is a sequence of data points indexed in time order, typically at uniform intervals such as daily, monthly, or quarterly. Unlike cross-sectional data—where each observation is independent—time series observations are inherently correlated because adjacent values share common influences. This temporal dependence is not a nuisance to be removed; it is the very structure that makes forecasting possible. Understanding a few foundational concepts allows analysts to break any business time series into interpretable components and select appropriate modeling strategies.

Trend

The long-run upward or downward movement of a series, reflecting structural forces such as population growth, technological change, or persistent shifts in consumer behavior.

Seasonality

Regular, repeating fluctuations tied to a known calendar period—weekly, monthly, or yearly. Retail sales spikes in Q4 and ice-cream demand peaks in summer are classic examples.

Cyclical Variation

Medium-to-long-term oscillations not tied to a fixed period, often driven by business cycles. Unlike seasonality, the length and amplitude of cycles vary unpredictably.

Irregular (Noise)

The residual variation that remains after trend, seasonality, and cycles are removed. It reflects random shocks—natural disasters, one-time promotions, or measurement error.

Stationarity

A statistical property indicating that the mean, variance, and autocorrelation structure of a series do not change over time. Most classical models require stationarity as a prerequisite.

✦

KEY TAKEAWAY

Think of a time series like a song. The trend is the melody gradually rising or falling in pitch. Seasonality is the recurring chorus that appears at predictable intervals. Cyclical variation is the tempo change between movements whose timing you cannot predict in advance. And noise is the ambient crowd murmur—random and unstructured. Decomposition is the equalizer board that lets you isolate each track.

SECTION 3

Visual Explanation — Decomposing a Time Series

The diagram below illustrates a classical additive decomposition of quarterly retail sales data over four years. The observed series (top panel) is separated into its constituent components: a linear upward trend, a repeating seasonal pattern, and an irregular residual. By visually inspecting each layer, an analyst can quickly judge whether a simple trend-plus-seasonality model might suffice or whether more sophisticated approaches are warranted.

The observed series (Y) equals the sum of the trend (T), seasonal component (S), and residual noise (ε). Notice how the seasonal peaks (Q2) repeat at fixed intervals while the trend slopes steadily upward.

In the diagram, the observed series (cyan line) clearly rises over time while oscillating. Stripping away the trend reveals a symmetric seasonal wave (pink), and once both are removed, only small, random residuals (amber) remain. When the residuals show no remaining pattern, the decomposition has captured the series' systematic structure. If instead the residuals still exhibit autocorrelation or heteroscedasticity, the analyst knows that a more flexible model—perhaps incorporating cyclical terms or time-varying variance—is needed.

SECTION 4

Mathematical Framework

The mathematics underlying time series analysis begins with two decomposition models—additive and multiplicative—and then extends into autocorrelation, which quantifies how strongly a series is related to its own lagged values. Mastering these building blocks is essential before applying models such as ARIMA or exponential smoothing.

ADDITIVE DECOMPOSITION

Yₜ = Tₜ + Sₜ + εₜ

Where Yₜ is the observed value at time t, Tₜ is the trend component, Sₜ is the seasonal component, and εₜ is the irregular residual. Use this model when seasonal fluctuations are roughly constant in absolute magnitude regardless of the level of the series.

MULTIPLICATIVE DECOMPOSITION

Yₜ = Tₜ × Sₜ × εₜ

In a multiplicative model, seasonal swings grow proportionally with the trend level. This is common in revenue data—a 10% holiday uplift on $1M in sales is far larger in absolute terms than a 10% uplift on $100K. Taking the natural log converts a multiplicative model into an additive one: ln(Yₜ) = ln(Tₜ) + ln(Sₜ) + ln(εₜ).

AUTOCORRELATION FUNCTION (ACF)

ρₖ = Cov(Yₜ, Yₜ₋ₖ) / Var(Yₜ)

The autocorrelation at lag k (ρₖ) measures the linear correlation between Yₜ and Yₜ₋ₖ. Values close to +1 or −1 indicate strong dependence; values near 0 indicate that observations separated by k periods are effectively uncorrelated. A slowly decaying ACF suggests a non-stationary trend; sharp spikes at regular lags suggest seasonality.

SIMPLE MOVING AVERAGE (TREND ESTIMATION)

T̂ₜ = (1/m) × Σᵢ₌₋ₖᵏ Yₜ₊ᵢ where m = 2k + 1

A centered simple moving average of order m smooths out short-term fluctuations to reveal the trend. For quarterly data, a 4-period moving average (often 2 × 4 to keep it centered) removes the seasonal component. Choosing m = period length ensures that exactly one full cycle is averaged away.

💡

Additive vs. Multiplicative: A Quick Decision Rule

Plot the series and examine the seasonal peaks and troughs. If the vertical distance between peaks and troughs stays roughly the same as the level rises, use the additive model. If the swings fan out (or narrow) proportionally, use the multiplicative model. When in doubt, take logs and test for constant variance.

SECTION 5

Stationarity & the ACF Plot

Most forecasting models, from ARIMA to exponential smoothing state-space forms, rest on the assumption of stationarity—the idea that the statistical properties of the series do not change over time. A stationary series has a constant mean, constant variance, and an autocovariance that depends only on the lag, not on the specific time at which it is computed. In practice, raw business data—quarterly revenue, daily stock prices, monthly website visits—is almost never stationary because it contains trends and seasonal patterns. The analyst's first task is therefore to transform the data into a stationary form, usually through differencing or log-transformation.

Left: a non-stationary ACF with slow decay indicating a trend or unit root. Right: after first-order differencing, the ACF shows a sharp cutoff into the significance band (shaded), confirming stationarity. The dashed horizontal lines mark the 95% confidence bounds.

In the left panel, the ACF bars decay slowly—lag 1 is very high, lag 2 slightly lower, and so on. This pattern is the hallmark of a non-stationary series driven by a trend or a unit root. After computing the first difference (Yₜ − Yₜ₋₁), the right panel shows that only lag 1 remains significant, with subsequent lags falling well within the 95% confidence band. This sharp cutoff confirms that the differenced series is stationary and suggests that a simple MA(1) process may describe the remaining autocorrelation—a direct clue for model selection in the Box–Jenkins framework.

ℹ

The Augmented Dickey–Fuller Test

While visual inspection of the ACF is informative, formal hypothesis testing via the Augmented Dickey–Fuller (ADF) test is standard practice. The null hypothesis is that the series contains a unit root (non-stationary). If the test statistic falls below the critical value, you reject the null and conclude the series is stationary. In most statistical software, a p-value < 0.05 supports stationarity.

SECTION 6

Worked Example — Decomposing Monthly Revenue

Suppose you are a business analyst at a retail company and you have 24 months of revenue data. You suspect an upward trend and a seasonal Q4 holiday spike. The goal is to perform an additive decomposition and estimate the deseasonalized (trend-adjusted) revenue for the most recent quarter.

Additive Decomposition of Monthly Revenue

1
Step 1 — Organize the Data

Arrange the 24 monthly observations in chronological order. Suppose the most recent 12 months of revenue (in $000s) are: 120, 115, 118, 130, 125, 122, 128, 140, 135, 130, 138, 155. The first 12 months show a similar pattern but shifted down by roughly $10K each month.
2
Step 2 — Estimate the Trend via Moving Average

Compute a 12-month centered moving average to remove the seasonal component. For each interior month t, calculate T̂ₜ = (1/12) × Σᵢ₌₋₅⁶ Yₜ₊ᵢ. Because a 12-month MA is even-ordered, apply a 2 × 12 centered average to align the estimate with integer months. For month 13 (the first month of Year 2), T̂₁₃ ≈ (sum of months 7–18) / 12 = (118 + 130 + 125 + 122 + 128 + 140 + 120 + 115 + 118 + 130 + 125 + 122) / 12.

T̂₁₃ ≈ $124.4K
3
Step 3 — Isolate the Seasonal + Residual

Subtract the trend from the observed value: Yₜ − T̂ₜ = Sₜ + εₜ. For month 13: 120 − 124.4 = −4.4. Repeat this for every month where the MA estimate exists.

Y₁₃ − T̂₁₃ = −4.4
4
Step 4 — Compute Seasonal Indices

Group the detrended values by calendar month and compute the average for each month across the available years. Suppose the average deviation for December (month 12) is +18.2 and for January (month 1) is −5.1. Adjust so all twelve indices sum to zero. These twelve values form the seasonal indices Ŝ₁, Ŝ₂, …, Ŝ₁₂.

Ŝ_Dec = +18.2, Ŝ_Jan = −5.1
5
Step 5 — Deseasonalize and Forecast

Subtract the seasonal index from the observed data to obtain the seasonally adjusted series: Y*ₜ = Yₜ − Ŝₜ. For the most recent December (Y₂₄ = 155), the deseasonalized value is 155 − 18.2 = 136.8. This tells management that underlying demand is about $136.8K after stripping out the holiday effect. To forecast January of Year 3, project the trend forward one month (say T̂₂₅ ≈ 138 via linear extrapolation) and add the January seasonal index: Ŷ₂₅ = 138 + (−5.1).

Ŷ₂₅ = $132.9K

SECTION 7

Strengths & Limitations of Classical Methods

Classical time series decomposition and moving-average techniques offer intuitive transparency and computational simplicity, making them excellent starting points for business forecasting. However, they come with well-known trade-offs that analysts should weigh before committing to a model for production use.

Dimension	Strengths	Limitations
Interpretability	Trend, seasonal, and residual components are directly visible and easily communicated to non-technical stakeholders.	Cannot capture complex, nonlinear interactions between components without extensions (e.g., STL decomposition).
Data Requirements	Work reasonably well with as few as 2–3 complete seasonal cycles (e.g., 2–3 years of monthly data).	Sensitive to outliers; a single anomalous quarter can distort the moving average and seasonal indices.
Forecast Horizon	Reliable for short-to-medium horizons (1–4 periods ahead) when the trend is linear and seasonality is stable.	Accuracy degrades rapidly for long horizons; assumes future patterns replicate the past exactly.
Adaptability	Simple to recalculate as new data arrives; no iterative optimization required.	Does not adapt to structural breaks (e.g., a pandemic, a new competitor) without manual intervention.
Uncertainty Quantification	Residual analysis provides a rough gauge of forecast variability.	No built-in prediction intervals; formal confidence bands require additional distributional assumptions.

Classical decomposition methods: strengths versus limitations

✦

KEY TAKEAWAY

Think of classical decomposition as a handheld compass rather than a GPS. It reliably points you in the right direction—identifying trend and seasonality—but it does not account for changing terrain, traffic conditions, or road closures. For high-stakes, long-horizon forecasts in dynamic environments, you will want to graduate to models like ARIMA, exponential smoothing (ETS), or machine learning approaches that can adapt to regime changes and provide probabilistic prediction intervals.

SECTION 8

Connection to Advanced Models

The concepts introduced in this lesson—decomposition, stationarity, and autocorrelation—are not stand-alone techniques but rather the diagnostic foundation upon which every advanced time series model is built. Understanding where classical methods end and modern methods begin helps you choose the right tool for each forecasting problem you encounter in a business setting.

Classical Concept	Advanced Extension
Moving-average trend estimation	Exponential Smoothing (ETS) — assigns exponentially decreasing weights to older observations, adapting the trend estimate more quickly to recent changes.
ACF / stationarity testing	ARIMA(p,d,q) — the Box–Jenkins methodology uses the ACF and PACF to select the order of autoregressive and moving-average terms after d-th order differencing.
Fixed seasonal indices	SARIMA / STL — Seasonal ARIMA adds seasonal AR and MA terms; STL decomposition (Seasonal and Trend decomposition using Loess) allows the seasonal pattern to evolve over time.
Residual variance assumed constant	GARCH — models time-varying volatility, essential for financial risk management where periods of high and low variance cluster together.
Linear trend extrapolation	Prophet / LSTM / Transformers — capture nonlinear trends, multiple seasonalities, holiday effects, and complex dependencies via additive regression (Prophet) or deep neural architectures.

From classical building blocks to advanced forecasting models

As you progress through your predictive modeling coursework, you will notice that every advanced method essentially automates or generalizes a step you have already learned here. ARIMA automates differencing and lag selection; exponential smoothing generalizes the moving average with adaptive weights; and machine learning models learn nonlinear versions of the trend and seasonal functions directly from data. The diagnostic tools—ACF plots, residual analysis, stationarity tests—remain indispensable regardless of model complexity, making this foundational knowledge deeply durable.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

A grocery chain observes that ice-cream sales increase every summer and decrease every winter, while overall sales have been climbing steadily for five years. Identify which component of a time series decomposition each pattern represents, and explain why it matters for inventory planning to distinguish between these two sources of variation.

Show answer

The summer/winter pattern is the seasonal component (S) because it repeats at a fixed annual interval. The five-year climb is the trend component (T). Distinguishing them matters because inventory planning requires the analyst to set a baseline order quantity that grows with the trend and then adjust upward or downward according to the seasonal index for each month. If the chain only looked at the raw totals, it might over-order in January (attributing the trend-driven increase to demand) or under-order next July (failing to anticipate the seasonal surge on top of the higher trend).

Question 2: BASIC CALCULATION

Given the following six quarterly revenue figures (in $M): 50, 55, 60, 65, 70, 75, compute the 4-quarter simple moving average centered on Q3 (the third observation, value = 60). Show your work.

Show answer

A 4-quarter moving average centered on Q3 uses the average of two consecutive 4-period means. First MA: (Q1 + Q2 + Q3 + Q4)/4 = (50 + 55 + 60 + 65)/4 = 57.5. Second MA: (Q2 + Q3 + Q4 + Q5)/4 = (55 + 60 + 65 + 70)/4 = 62.5. The centered moving average at Q3 is (57.5 + 62.5)/2 = 60.0. In this case the centered MA equals the original value, which makes sense because the series is perfectly linear—there is no seasonality to smooth out.

Question 3: INTERMEDIATE

An analyst computes the ACF of a monthly sales series and finds ρ₁ = 0.92, ρ₂ = 0.85, ρ₃ = 0.78, and ρ₁₂ = 0.60 with a slow, linear decay across all lags. After taking first differences, the ACF shows ρ₁ = 0.35, ρ₂ = 0.05, ρ₃ = −0.02, and all higher lags within the 95% confidence band. (a) What does the original ACF pattern indicate? (b) How many times should the series be differenced? (c) Based on the differenced ACF, what ARIMA(p,d,q) model would you tentatively propose?

Show answer

(a) The slow, linear decay of the original ACF indicates non-stationarity due to a trend or unit root. (b) One round of differencing (d = 1) is sufficient because the differenced ACF drops into the confidence band quickly. (c) After differencing, only ρ₁ = 0.35 is significant. A single significant ACF spike followed by a cutoff is the signature of a moving-average process of order 1. The tentative model is therefore ARIMA(0, 1, 1). One would confirm this by also examining the PACF and comparing information criteria (AIC/BIC) with competing specifications.

Question 4: APPLIED

A SaaS company's monthly recurring revenue (MRR) has grown from $200K to $800K over three years. The CFO notes that revenue dips every February (contract resets) and spikes every November (annual renewals). When the analyst plots the series, the February dips grow from −$8K in Year 1 to −$25K in Year 3, and the November spikes grow from +$15K to +$50K. Should the analyst use an additive or multiplicative decomposition model? Justify your answer quantitatively and explain one consequence of choosing the wrong model.

Show answer

The analyst should use a multiplicative decomposition. Quantitatively, the February dip as a proportion of the level is roughly −8/200 = −4% in Year 1 and −25/800 ≈ −3.1% in Year 3—fairly stable percentages. In absolute terms, however, the dip more than tripled. Similarly, November spikes are +15/200 = 7.5% and +50/800 = 6.25%—roughly proportional to the level. An additive model would assume constant absolute seasonal swings, so it would understate November's revenue at high MRR levels and overstate February's dip at low MRR levels, leading to biased forecasts. Using the multiplicative model (or equivalently, log-transforming and applying additive decomposition) ensures the seasonal indices scale with the trend.

Question 5: CRITICAL THINKING

A ride-sharing company experienced a dramatic ridership drop in March 2020 due to the COVID-19 pandemic, followed by a gradual but incomplete recovery. The company's forecasting team uses an ARIMA model trained on 2017–2019 data. (a) Explain why the model's residuals from March 2020 onward would likely violate stationarity assumptions. (b) Propose two strategies the team could employ to adapt their time series framework to this structural break. (c) Discuss the trade-off each strategy involves.

Show answer

(a) The ARIMA model was estimated on pre-pandemic data where the mean level and variance were stable. The pandemic introduced a structural break—a permanent shift in the mean level—that the model cannot capture with its fixed parameters. Residuals from March 2020 onward would show persistent negative bias (underprediction) and autocorrelated errors, violating the stationarity assumption of the residual process. (b) Strategy 1: re-estimate the model using only post-break data (e.g., June 2020 onward), discarding pre-pandemic observations. Strategy 2: include a structural-break indicator variable (intervention analysis) or use a piecewise model that allows the trend intercept and possibly the slope to shift at a known break date. (c) Strategy 1 trades history for relevance—using only post-break data reduces sample size, increasing parameter uncertainty, but ensures the model reflects the new regime. Strategy 2 preserves pre-pandemic data and statistical power but assumes the seasonal pattern and autocorrelation structure did not fundamentally change, which may not hold if consumer behavior shifted permanently.

SUMMARY

Time Series Basics — Summary

Summary

A time series is a sequence of observations indexed by time, and its power for business forecasting lies in the temporal dependence among consecutive values. Every series can be decomposed into a trend (long-run direction), seasonality (fixed-period oscillations), cyclical variation (irregular oscillations not tied to a fixed period), and noise. The decomposition may be additive (Yₜ = Tₜ + Sₜ + εₜ) when seasonal amplitude is constant, or multiplicative (Yₜ = Tₜ × Sₜ × εₜ) when seasonal swings scale with the level.

Before fitting any forecasting model, analysts must verify stationarity—the requirement that the series' mean, variance, and autocorrelation structure remain constant over time. The autocorrelation function (ACF) is the primary diagnostic tool: a slowly decaying ACF signals non-stationarity, typically resolved through differencing. These foundational ideas—decomposition, stationarity, and autocorrelation—underpin every advanced model from ARIMA and exponential smoothing to modern deep-learning architectures, making them essential knowledge for any business analytics professional.