Introduction to Time Series

Introduction to Time Series Understanding Time Series A time series is a sequence of observations measured at successive points in time, typically at regular intervals such as daily, monthly, or yearly. Unlike a simple snapshot of data, a time series allows you to observe how a variable changes over time rather than just what its value is at any single moment. The crucial insight is that the order of observations matters. This is fundamentally different from most data you might have encountered before. In a time series, observations that are close together in time tend to be related to each other. For example, today's temperature is likely similar to tomorrow's temperature, but very different from the temperature six months from now. This property—called temporal dependence—is what makes time series analysis special. Why Order Matters: Time Series vs. Cross-Sectional Data Consider two types of data: Cross-sectional data captures many units at a single point in time. For instance, the heights of students in a classroom recorded on Monday morning. The order you list the students doesn't matter—height doesn't depend on whether you measure student A before or after student B. Time series data, by contrast, tracks one or more variables over time. The temperature readings in a city measured every day for a year form a time series. Here, order is everything: you can't scramble the readings randomly without losing important information. Components of a Time Series Most time series can be decomposed into distinct components, each capturing different patterns in the data: Trend represents the long-term direction of movement in the data. This could be a steady increase (like population growth in a country over decades) or a decrease (like declining usage of a technology). The trend captures the "big picture" direction over an extended period. Seasonality refers to regular, repeating patterns that occur within a fixed period. These cycles return consistently—perhaps annually, quarterly, or even daily. For example, electricity demand is typically higher during summer months and lower during winter, and this pattern repeats every year. Retail sales spike around the winter holidays. These predictable patterns tied to calendar periods define seasonality. Cyclical fluctuations are longer-term waves that are not tied to a fixed calendar period. Business cycles are the classic example: economies expand and contract, but the length of each cycle varies. A cycle might last several years, and you can't predict exactly when the next peak or trough will occur just from knowing the date. Irregular or noise component captures random variations that don't fit the other patterns. These are sudden shocks or unexplained variations—like a temporary factory shutdown, an unusual weather event, or measurement error. This component represents what we can't explain with trend, seasonality, or cyclical patterns. The image above shows a real example: tuberculosis deaths over time (top), with the series decomposed into trend (showing the overall decline), yearly changes (seasonality), and percent change. Tools for Time Series Analysis Visual Exploration The first step in analyzing any time series is to plot it. A visualization reveals trends, seasonal patterns, outliers, and structural breaks at a glance. Looking at the raw data helps you ask better questions and understand what analysis techniques might be appropriate. Smoothing Techniques Moving averages are a simple but powerful tool. The basic idea is to replace each observation with the average of that observation and its neighbors. For example, a 3-point moving average replaces each value with the mean of that value and the two adjacent values. This smoothing reduces noise and makes underlying patterns clearer, like wiping fog off a window to see the landscape beneath. Decomposition Methods Decomposition formally separates a time series into its constituent components: trend, seasonal, and irregular. This technique gives you estimates of each component, allowing you to study them separately. If you have a time series $yt$ at time $t$, decomposition might express it as: $$yt = \text{Trend}t + \text{Seasonal}t + \text{Irregular}t$$ This is incredibly useful for understanding which patterns dominate your data and for removing seasonality if you want to focus on trend. Fundamental Forecasting Models The Naïve Forecast The simplest possible forecasting model is the naïve forecast, which predicts that the next value will be exactly the same as the last observed value: $$\hat{y}{t+1} = yt$$ While simplistic, the naïve forecast often serves as a useful benchmark. If your sophisticated model doesn't beat the naïve forecast, something is wrong. Simple Exponential Smoothing Simple exponential smoothing is a more refined approach that updates forecasts based on the current observation and the previous forecast. The updating formula is: $$St = \alpha Xt + (1-\alpha) S{t-1}$$ where: $St$ is the smoothed value (or forecast) at time $t$ $Xt$ is the actual observation at time $t$ $\alpha$ is the smoothing parameter, a number between 0 and 1 $S{t-1}$ is the smoothed value from the previous time period The smoothing parameter $\alpha$ controls how much weight you give to recent observations versus past history. If $\alpha = 1$, you only look at the current observation. If $\alpha = 0$, you ignore the current observation entirely and stick with past forecasts. In practice, $\alpha$ values between 0.1 and 0.3 often work well, balancing responsiveness to new information with stability. Autoregressive Integrated Moving Average (ARIMA) The ARIMA model is a powerful and flexible framework that combines three components to capture different aspects of time series behavior: Autoregressive (AR) terms use past values of the series to forecast future values, based on the idea that a series has momentum or persistence. Integrated (I) terms involve differencing—computing changes between consecutive observations. This transformation helps stabilize a series that has a trend, converting it into a stationary series (one with constant mean and variance over time). Moving Average (MA) terms model the dependency of current values on past forecast errors, capturing how recent shocks propagate through the system. ARIMA models are denoted as ARIMA$(p,d,q)$, where $p$ is the number of autoregressive terms, $d$ is the degree of differencing, and $q$ is the number of moving average terms. This framework can represent a wide variety of time series patterns, from simple random walks to complex cyclical behavior. <extrainfo> Cyclical Components (More Detail) While cyclical fluctuations are mathematically similar to seasonal patterns, they differ fundamentally in one key way: cyclical patterns have no fixed period. Unlike seasonality, which repeats on a predictable calendar schedule, cycles can vary in length. This makes cyclical patterns harder to model and forecast precisely. In many analyses, cyclical components are included in either the trend or irregular component rather than treated separately. Transformation Methods When working with time series, you sometimes need to transform the data before analysis. For example, if variance increases over time, taking logarithms or square roots can stabilize it. The image below shows how different transformations affect data shape: Each transformation curve has different properties: linear transformations do nothing, square root and logarithmic transformations compress large values more than small ones, while exponential transformations do the opposite. </extrainfo>