Every morning, the National Weather Service issues a temperature forecast for the next ten days. Every quarter, the Federal Reserve forecasts GDP growth. Every week, epidemiologists forecast the trajectory of a flu outbreak. All of these predictions share a common challenge: the data arrives in sequence over time, and past values carry information about future ones. The mathematical tools for extracting that information — and building reliable forecasts — are called time series analysis.
What Makes Time Series Data Special
Ordinary statistical data treats observations as independent — one person's test score doesn't affect another's. Time series data is different: today's temperature is correlated with yesterday's, and yesterday's with the day before. This correlation structure isn't a nuisance to be ignored — it's the signal to be exploited. A good time series model captures how the past predicts the future.
Every time series can be decomposed into three components: trend (the long-run direction — rising, falling, or flat), seasonality (regular repeating cycles — weekly, monthly, or yearly patterns), and noise (random fluctuations that no model can predict). Separating these three components is the first step in any serious time series analysis.
Autocorrelation: How Much Does the Past Matter?
Autocorrelation measures how strongly a time series is correlated with its own past values. The autocorrelation at lag 1 measures the correlation between each value and the value one time step before it. At lag 7, it measures the correlation with the value 7 steps back.
Plotting autocorrelation across all lags reveals the structure of a time series. If autocorrelation drops quickly to zero, each value is only weakly connected to the past — short-term forecasting is feasible but long-term forecasting is hard. If autocorrelation stays high for many lags, the series has strong persistence — yesterday (or last week or last year) remains informative for a long time.
ARIMA Models: A Standard Forecasting Framework
The most widely used time series forecasting framework is ARIMA — Autoregressive Integrated Moving Average. Breaking it down: "Autoregressive" means predicting the current value from past values. "Moving Average" means accounting for past prediction errors. "Integrated" means differencing the series (subtracting each value from the previous) to remove trends and make the series stationary — fluctuating around a stable mean.
The Weather Service's ten-day forecast uses a much more complex version of this principle — dozens of variables, spatial correlations, and physics-based constraints — but the mathematical core is the same: quantify how much the past predicts the future and extrapolate forward.
Flu Forecasting
The CDC tracks weekly flu hospitalizations. Time series analysis reveals a strong seasonal pattern — flu peaks every winter — plus shorter-term autocorrelation (this week's case count predicts next week's). Combining seasonal decomposition and ARIMA modeling, epidemiologists produce forecasts 4 to 8 weeks ahead that guide hospital staffing and vaccine distribution decisions. During the COVID-19 pandemic, the same methods — adapted for a non-seasonal pattern with rapidly changing dynamics — were the primary forecasting tools used by public health agencies worldwide.
Conclusion
Time series analysis extracts the temporal structure from sequential data — trends, seasonal patterns, and autocorrelation — and uses that structure to forecast the future. The key insight is that time-ordered data isn't just data: the order carries information. Autocorrelation quantifies how much, and ARIMA models use that information to make principled predictions. From weather forecasts to economic projections to epidemic tracking, this mathematical framework is how we transform the patterns of the past into reliable estimates of what comes next.