Data in which the order of observations carries meaning, so that the sequence of elements conveys information beyond what any individual element contains alone. Sequential data includes time series measurements, text as ordered token sequences, user behavioral logs as ordered event streams, and any dataset where reordering the observations would change the information they convey about the underlying process.
Also known as time-ordered data, temporal data, ordered data streams
Sequential data has a defined ordering that is inherent to the data-generating process rather than arbitrary. Text is sequential because the order of words determines meaning; swapping two words in a sentence typically changes or destroys the sentence’s meaning. Time series data is sequential because earlier values are the causal antecedents of later values, and the temporal ordering reflects the direction of causality. User behavioral logs are sequential because the sequence of actions within a session captures the user’s intent progression: viewing a product page before a cart add is a different behavioral signal than cart-adding before viewing the product page, even though both events occurred in the same session.
The sequential nature of data has direct implications for how it should be processed, split for training and validation, and evaluated. For non-sequential tabular data, random shuffling and splitting into train and validation sets is appropriate. For sequential data, this is invalid: randomly shuffling time series data before splitting destroys the temporal structure and produces a validation set that contains future observations from the training period, which creates data leakage. Sequential data must be split chronologically: earlier observations for training, later observations for validation and testing. Models should be evaluated on their ability to predict future observations given only past observations, not their ability to interpolate within the historical data.
Sequential data has different statistical properties than cross-sectional data. Autocorrelation, where an observation at time t is correlated with observations at times t-1, t-2, and so on, violates the independence assumption that most standard statistical tests require. Non-stationarity, where the statistical properties of the data change over time (trending upward, shifting in mean or variance), causes models trained on historical data to systematically mis-predict future values that fall outside the historical range. Stationarity testing and differencing to remove trends are preprocessing steps specific to sequential data that are unnecessary for cross-sectional data.
A working ad agency building models on campaign performance data, sales forecasts, customer behavioral sequences, or any other time-ordered data source needs to handle sequential data differently from cross-sectional data at every stage of the modeling pipeline: during preprocessing, during train-validation splitting, during feature engineering, and during model evaluation. Applying cross-sectional modeling practices to sequential data introduces data leakage and produces inflated validation metrics that do not reflect the model’s actual ability to predict future outcomes, which is the only prediction task that matters in deployment.
Chronological train-validation splitting is required for any model that will be used to predict future values from past observations. A media performance prediction model split randomly into 70% train and 30% validation will include future observations in the training set and past observations in the validation set, training the model on data from the same time period it is being validated on. This data leakage inflates validation metrics because the model has partially seen the future. The correct split holds out the most recent time window as validation (for example, the most recent 3 months as the validation set and all prior data as training), testing whether the model generalizes to a time period it has never seen rather than interpolating within the historical data it was trained on.
Feature engineering for sequential data must use only information available at the time of prediction to avoid lookahead bias. A feature engineering step that computes a 4-week rolling average of campaign spend as a predictor must use only the 4 weeks prior to the prediction date, not a centered window that includes future observations. Using a centered window that incorporates future data leaks information from the future into the training features, producing a model that appears accurate in backtesting but fails in forward-looking deployment because the future values it was trained on are not available when the model is actually deployed. All rolling statistics, lags, and window-based features for sequential data must be computed using only backward-looking windows anchored to the prediction date.
Autocorrelation in sequential marketing data means that current performance is partially predictable from recent history, and ignoring this in model design underperforms even simple baselines. A model that predicts next-week campaign CTR using only the input features without including the prior week’s CTR as a predictor misses the strongest single predictor for many campaign performance metrics: their own recent history. A campaign that achieved 2.1% CTR last week is much more likely to achieve 1.8 to 2.4% CTR this week than a campaign with no prior CTR history. Including appropriate lag features (prior week value, prior 4-week average, prior year same-week value) provides the model with the autocorrelative structure of the data, substantially improving forecast accuracy over feature-only models that ignore sequential dependencies.
An agency is building a weekly budget pacing alert system for a portfolio of 68 active client campaigns, predicting end-of-month spend versus monthly budget to identify campaigns on track to over- or underpace by more than 10% before the issue becomes unsalvageable. Initial model development uses a random forest on daily spend and performance features, with train-validation split chosen randomly. Validation MAPE: 6.2% for the random split. After agency review identifies the split as invalid for sequential data (the validation set includes historical periods from the same months as the training set), the agency retrains with a proper chronological split (most recent 2 months as validation). Corrected validation MAPE: 14.8%, revealing that the original 6.2% metric was inflated by data leakage. The agency adds lag features: prior week’s daily spend, rolling 4-week average daily spend, days elapsed in month, and prior month’s final spend versus budget ratio. The lag-augmented model with chronological split achieves validation MAPE of 9.3%, a genuine improvement over the 14.8% baseline and better than the naive “current pacing rate times remaining days” heuristic the account team was using (MAPE: 17.1%). The system correctly flags 73% of campaigns that ultimately overpaced or underpaced by more than 10%, with a 14% false positive rate (campaigns flagged that finish within 10% of budget). The alerts are issued by Wednesday of each week, giving account managers 2 to 3 days to adjust campaign settings before weekend spend, when intervention is less operationally feasible.
The generative AI foundations module covers sequential data handling including chronological splitting, lookahead bias prevention, autocorrelation and stationarity, and feature engineering practices that correctly respect the temporal structure of marketing performance data.