A financial time series is a sequence of financial or market observations indexed through time. Examples include asset prices, returns, exchange rates, yields, trading volume, volatility estimates and portfolio values. The defining feature is not simply that there are many observations; it is that their order in time carries information.
Time order changes the analysis
In an ordinary cross-sectional dataset, rearranging rows may leave the statistical question unchanged. In a time series, rearranging observations destroys the sequence that connects one period to the next.
This matters because financial observations can depend on earlier conditions. Volatility may cluster, liquidity can change by session, relationships can drift and the information available today is not the same as the information available later.
Prices and returns answer different questions
A price series describes the level at which an asset was valued or traded. A return series describes the change in value over an interval. Research often works with returns because they make changes across assets and periods easier to compare, but the transformation must match the question.
Simple Returns vs Log Returns explains the arithmetic and compounding differences. A price series should not be treated as interchangeable with its return series simply because both come from the same asset.
Sampling frequency is part of the definition
Daily closes, hourly bars, one-minute observations and individual transactions describe different information sets. Higher frequency can reveal intraday structure, but it can also introduce bid-ask effects, irregular spacing, market microstructure noise and larger data volumes.
State the sampling rule, timezone, session boundary and treatment of missing observations. A timestamp without a clear market convention can make apparently precise data ambiguous.
Dependence can occur in levels, returns or volatility
Autocorrelation measures dependence between a series and lagged versions of itself. Weak autocorrelation in raw returns does not imply that every transformation is independent; absolute or squared returns, for example, can display different dependence patterns.
Dependence is evidence about a statistical relationship. It is not automatically a tradable signal after costs and execution.
Statistical properties can change over time
Many time-series models rely on some form of stability in the data-generating process. Stationarity formalizes versions of that requirement. Financial data can experience structural breaks, volatility changes and regime shifts, so stability should be tested and interpreted for the particular variable and horizon rather than assumed universally.
Volatility is itself a time-series problem
Periods of large and small market moves often arrive in clusters. Volatility Modelling studies how variability changes over time, while GARCH Models cover one specific family of conditional-variance models.
A volatility estimate describes uncertainty in magnitude, not the direction of the next return.
Data availability must be historical, not reconstructed with hindsight
A research dataset should distinguish the event time from the time the information became usable. Revised economic data, corrected prices, corporate actions and later-added classifications can create look-ahead bias if a model is allowed to use information that was unavailable at the decision date.
This is why Time-Series Cross-Validation preserves temporal order during evaluation.
Common data problems can dominate the model
- missing or duplicated timestamps;
- incorrect timezone or daylight-saving treatment;
- unadjusted corporate actions where adjusted data are required;
- survivorship or selection bias in the asset universe;
- future information leaking into features;
- mixing frequencies or sessions without an explicit rule.
These are research-design problems, not problems that a more complex algorithm automatically repairs.
Use the series that matches the decision
A trading-entry study may need intraday observations. A portfolio-allocation study may need lower-frequency total returns. A risk model may need a return series plus volatility and correlation information. The correct time series is therefore defined by the decision, not by the largest dataset available.
Financial time-series analysis begins by preserving what happened, when it happened and what could genuinely have been known at that time. Once that foundation is credible, statistical models can be used without confusing hindsight with evidence.