Time-series analysis studies observations in temporal order to describe their behaviour, model dependence and assess what can be inferred or forecast. In quantitative finance, the ordering matters twice: it defines which observations are lags, and it determines what an analyst could have known when making a decision. A convincing fit can be unusable if it depends on tomorrow's information.
A time series is the ordered data; time-series analysis is the work performed on it. Prices, returns, interest rates, trading volumes and realised volatility can each be analysed, but they answer different questions.
Start with the quantity and the clock
Suppose a synthetic asset closes at USD 100, 110 and 100 per unit on trading days 0, 1 and 2. Assume no cash distributions, transaction costs or external flows. Its one-period simple returns are
Holding one unit across both periods produces a compounded price return of
Adding the two simple returns would give a different answer because their starting prices differ. This tiny sequence illustrates measurement and timing; it cannot establish a reliable forecasting model.
Retrospective view of synthetic, equally spaced trading-day closing prices in USD per unit. Filled points are available after the day 1 close; the hollow day 2 point and dashed segment are revealed later. The plotted path is not a forecast.
Shuffling the price observations can change which prices are adjacent and therefore change the computed returns and lags. Shuffling already-computed returns preserves their compounded product, but can change the path and a time-dependent signal. Neither operation preserves the original forecasting problem.
The clock also needs a definition. “One lag” might mean one trading day, one calendar month or one transaction. Missing observations and irregular timestamps require explicit handling; an absent quote is not automatically a zero return.
From a plot to a model
The analysis should follow the financial question rather than begin with a favourite model.
| Task | What the analyst asks |
|---|---|
| Describe | Are there trends, unusual observations, changing dispersion or autocorrelation? |
| Specify | Is the target a price level, a return, or future risk? Is stationarity plausible for the chosen transformation and window? |
| Estimate | Which parameters can be supported by the available history without fitting incidental noise? |
| Diagnose | Do residuals retain systematic patterns? Does their variability change over time? |
| Evaluate | Does the method work on later observations, at the required horizon, against a relevant baseline? |
Mean and risk dynamics need separate attention. Returns can have little linear autocorrelation while their magnitudes exhibit volatility clustering. A model of the conditional mean alone can miss this risk dependence. Conversely, passing residual autocorrelation checks does not prove independence or guarantee accurate future forecasts.
For the formal modelling loop, Box–Jenkins: identification, estimation and diagnostic checking develops ARMA model selection and residual assessment. The full lesson requires membership; its public preview shows the scope.
Evaluate from the information available then
After the day 1 close in the example, a forecast for day 2 may use the first two prices. The last price is available only afterward, when the forecast can be scored. A retrospective chart shows both sides of that boundary; a forecasting procedure must respect it.
A forward evaluation repeats this separation: fit on an earlier window, forecast later observations, then advance the forecast origin. Fit data transformations and select model settings using only information permitted by that evaluation design. Revised economic data require the vintage available at the historical decision time. Otherwise look-ahead bias can enter before the model is even fitted.
Choose the baseline and loss measure for the task: a last-observed-price forecast is a useful price benchmark, while a risk forecast needs an assessment of its uncertainty predictions. A good statistical score also does not establish a profitable strategy; a trading application needs separate backtesting with execution and costs.
Sources
- Hyndman, R. J. and Athanasopoulos, G. Forecasting: Principles and Practice, 3rd ed., §1.4: Forecasting data and methods.
- Hyndman and Athanasopoulos, Chapter 5: The forecaster's toolbox, especially residual diagnostics and time-series cross-validation.
- Brooks, C. (2008). Introductory Econometrics for Finance, 2nd ed., Cambridge University Press, §§5.1–5.2 and 5.7, pp. 206–209 and 230–233.