Glossary¶

Key terms used across Yohou documentation.

Forecasting¶

Forecasting horizon: The number of future timesteps to predict. Specified at fit() time and optionally overridden at predict() time.
Observation horizon: The number of recent time steps a stateful component must retain in memory to produce output. A LagTransformer with lags=[1, 7] has an observation horizon of 7. Forecasters derive theirs from the maximum across their transformers. See Core Concepts.
Memory buffer: The internal store of recent rows (up to observation_horizon in length) that a stateful component maintains. Updated by observe() and trimmed by rewind().
Observe: Updating a fitted forecaster's or transformer's internal buffers with newly arrived data, without refitting. Called via observe(), typically as part of the composite observe_predict() during rolling evaluation.
Rewind: Resetting observation buffers to the last observation_horizon rows. Called automatically after observe() to keep memory bounded.
Composite method: A method that combines two operations in sequence: observe_predict, observe_predict_interval, observe_predict_class_proba, observe_transform, and rewind_transform. These are routed together through metadata routing. See Core Concepts.
Point forecast: A single numeric prediction per timestep, produced by predict().
Interval forecast: A pair of bounds \([L, U]\) per timestep and coverage rate, produced by predict_interval(). Designed so that the true value falls within the interval at least \(\alpha\)% of the time.
Class-probability forecast: A probability distribution over categorical classes per timestep, produced by predict_class_proba(). Each row sums to 1. See Class-Probability Forecasting.
Coverage rate: The target probability \(\alpha\) that an interval forecast should contain the true value. Common values: 0.90, 0.95.
Calibration: The degree to which predicted probabilities or intervals match observed frequencies. A well-calibrated 90% interval contains the true value 90% of the time.
Recursive prediction: How a forecaster predicts beyond its training horizon by feeding its own predictions back as inputs and applying itself iteratively. See Reduction Forecasting.
Error accumulation: The compounding of prediction errors when recursive prediction feeds earlier predictions back as inputs. The direct reduction strategy avoids this by fitting independent models per step. See Reduction Forecasting.
Vintage: A single forecast origin: the point in time at which the forecaster last observed data before predicting. Each call to observe_predict during rolling evaluation produces one vintage. See Core Concepts.
vintage_time: The column in a prediction DataFrame that records the last observed timestamp for each vintage. Used by scorers to support vintagewise aggregation and by plotting functions like plot_score_per_vintage.
Step: A single position within the forecasting horizon (1-based). Step 1 is the one-step-ahead prediction; step \(H\) is the \(H\)-step-ahead prediction. Used by scorers with aggregation_method="stepwise" and by plot_score_per_step.
Stride: The number of time steps the cross-validation window advances between folds. A stride equal to test_size produces non-overlapping test windows; a smaller stride produces overlapping windows for finer-grained evaluation. Also used in observe_predict to control how many time steps to advance between successive prediction vintages.
Rolling evaluation: A procedure for assessing forecaster performance over time. The forecaster is trained once, then repeatedly observes new data and predicts the next horizon, producing one vintage per iteration. Also called walk-forward validation. See Model Selection.

Data¶

Time column contract: Every DataFrame in Yohou must have a "time" column containing datetime values. This column is preserved through all transformations. See Core Concepts.
Polars-native: Yohou uses Polars DataFrames as its internal data representation throughout the entire pipeline (fitting, transforming, predicting, scoring). All user-facing inputs and outputs are Polars DataFrames.
Panel data: Multiple related time series handled together. Groups are identified by a prefix in column names using the __ separator. See Panel Data.
Group prefix (__): The double-underscore separator between a panel group name and a column name. For example, store_A__sales identifies column sales in group store_A.
Panel strategy: Controls how a forecaster handles panel data. "global" (default) fits one model with per-group transformer state. "multivariate" skips panel detection and treats __-prefixed columns as ordinary wide-format columns.
Exogenous features: Additional input columns (beyond the target) that may improve forecasts. Passed as the X parameter in fit() and predict(). See Exogenous Features.
Known-future features: Exogenous features whose values are available for the prediction horizon at forecast time (e.g., holidays, scheduled events). Passed in the X argument of predict(). See Exogenous Features.
Step-indexed columns: Feature columns named with a _step_h suffix that carry different values for each forecasting step. Used to provide step-specific exogenous information to direct or dir-rec reduction strategies. See Exogenous Features.
Univariate: A single target column (plus time). The simplest forecasting setting, where the model predicts one variable using only its own past values and optionally exogenous features.
Multivariate: Multiple target columns forecasted simultaneously. Each column receives its own predictions, but all columns share the same time index and can influence each other through shared features or model structure.

Modeling¶

Tabularization: The process of converting a time series into a flat feature matrix that a standard sklearn estimator can consume. Lag features, rolling statistics, and calendar attributes are typical columns in the tabularized output.
Reduction strategy: Converting a time series forecasting problem into a tabular supervised learning problem. Three strategies control how multi-step horizons are handled: multi-output fits one model that predicts all steps simultaneously, direct fits one independent model per step, and dir-rec fits models sequentially so each step can use predictions from earlier steps. See Reduction Forecasting.
Step feature alignment: Controls which step-indexed columns each direct estimator receives in a direct or dir-rec reduction strategy. Options: "all" (every step column), "matched" (only the column matching the current step), and "cumulative" (columns for the current step and all preceding steps).
Pipeline: A sequence of transformers followed by a forecaster, executed in order. Yohou pipelines propagate observe() and rewind() calls through each step and accumulate observation horizons. See Feature Pipelines.
Target transformer: A BaseTransformer applied to the target series y before tabularization. Used for operations like differencing or scaling that should be inverted after prediction.
Feature transformer: A BaseTransformer applied to the feature matrix X before tabularization. Used for creating lag features, rolling statistics, or other derived inputs.
Stateful transformer: A BaseTransformer that maintains an internal observation window of recent values and updates it during observe(). Stateful transformers have a non-zero observation_horizon because they need to remember past values (e.g., lag features, rolling statistics) to transform new data. See Preprocessing for details on how observation horizons propagate through pipelines.
Stateless transformer: A BaseTransformer whose output depends only on its fitted parameters and the current input, with no dependence on prior observations. Has an observation_horizon of 0. Examples include scaling, log transforms, and calendar feature extraction.
Ensemble: Combining predictions from multiple forecasters to reduce variance. Implemented via the VotingPointForecaster, VotingIntervalForecaster, and VotingClassProbaForecaster classes. See Ensemble Forecasting.
Forecaster composition: Building complex forecasting workflows by combining simpler components. Includes pipelines, column forecasters (different forecasters per column), decomposition pipelines (trend, seasonality, residual), and forecasted-feature forecasters (two-stage prediction of exogenous inputs). See Forecaster Composition.
Decomposition: Separating a time series into additive components (trend, seasonality, residual) for separate modeling, implemented in DecompositionPipeline. See Stationarity.
Variance stabilization: Transforming a time series so that its error variance is approximately constant over time. Common transforms include logarithms, Box-Cox, and inverse hyperbolic sine. Useful when error magnitude scales with the level of the series. See Stationarity.
Stationarity: A time series is stationary when its statistical properties (mean, variance) do not change over time. Many forecasting methods assume or benefit from stationarity, and Yohou provides transforms like SeasonalDifferencing and DecompositionPipeline to achieve it before modeling. See Stationarity.
Conformal prediction: A distribution-free method for constructing prediction intervals with finite-sample coverage guarantees, implemented in SplitConformalForecaster. See Interval Forecasting.
Conformity score: The residual measurement computed on a held-out calibration set during conformal prediction. The quantile of conformity scores determines the width of prediction intervals. Yohou provides signed, absolute, gamma, and quantile variants.
Calibration set: A portion of the training data held out from model fitting and used exclusively to compute conformity scores for interval construction. The quality of conformal intervals depends on the calibration set being representative of future data.
Similarity measure: A function that weights conformity scores based on the similarity between the current prediction context and past calibration contexts. Produces locally adaptive intervals that are wider in unfamiliar regimes and narrower in well-represented ones. See Interval Forecasting.
Metadata routing: The mechanism by which sample-level metadata (e.g., time_weight, step_weight, vintage_weight) flows from a composite estimator (pipeline, search, ensemble) down to its child components. Yohou enables scikit-learn's metadata routing globally and registers seven additional routable methods. See Metadata Routing.

Evaluation¶

Scorer: An object that computes a numeric quality measure from predicted and actual values. Yohou provides point scorers (e.g., MAE, RMSE, MASE), interval scorers (e.g., IntervalScore, EmpiricalCoverage), and class-probability scorers (e.g., LogLoss, BrierScore). All scorers share a common score() interface. See Forecast Accuracy.
Proper scoring rule: A metric that is uniquely minimized (or maximized) when the predicted distribution matches the true distribution. LogLoss and BrierScore are proper scoring rules; Accuracy is not.
Aggregation method: Controls which dimensions a scorer collapses when computing its result. Options: "timewise" (across time), "componentwise" (across multivariate columns), "groupwise" (across panel groups), "stepwise" (across forecasting steps), "vintagewise" (across vintages), "coveragewise" (across coverage rates, intervals only), and "all" (collapse every dimension to a single scalar). See Forecast Accuracy.
Forecast error: The difference between a predicted value and the corresponding actual value, computed on out-of-sample data. Distinct from a residual, which is computed on training data. See Forecast Accuracy.
Scale-dependent metric: A metric whose value depends on the scale of the data (e.g., MAE, RMSE). Cannot be compared directly across series with different units or magnitudes.
Scale-independent metric: A metric that normalizes errors so they can be compared across series of different scales (e.g., MAPE, MASE).
Cross-validation: Evaluating model performance by repeatedly splitting data into training and test sets. Time series cross-validation uses temporal splits to prevent data leakage.
Temporal split: A train/test split that respects time ordering: training data always precedes test data chronologically. This prevents future information from leaking into training.
Expanding window splitter: A cross-validation splitter where each fold grows the training window while keeping the test window a fixed size. The training set includes all data from the beginning of the series up to each cutoff point. See Model Selection.
Sliding window splitter: A cross-validation splitter that maintains a fixed-size training window that slides forward in time. Older data is dropped from training as newer data is added, making it better suited for data with concept drift or regime changes. See Model Selection.
Concept drift: A shift in the statistical properties of the data over time, causing a model trained on older data to become less accurate. Sliding window splitters mitigate this by limiting training to recent observations.
Leakage: Using information from the future (test period) during training, leading to artificially optimistic performance estimates. In time series, leakage commonly occurs from improper cross-validation splits or applying global normalization before splitting. See Model Selection for how Yohou's splitters prevent temporal leakage.
Time weighting: Applying non-uniform weights to observations or errors so that specific time periods carry more or less influence. Yohou supports three weight types: time_weight (per-timestep), step_weight (per-forecasting-step), and vintage_weight (per-forecast-origin). See Weighting for formats and normalization.
time_weight: A per-timestep weight. In training, it controls the emphasis on individual observations. In scoring, it controls how much each time point contributes to the aggregated metric. Passed as a column or Series alongside the data.
Step weight: A weight applied per forecasting step (1-step-ahead, 2-step-ahead, etc.) during scoring. Controls how much each lead time contributes to the aggregated score. Only available in scoring, not training.
Vintage weight: A weight applied per vintage (forecast origin date). In training, controls per-observation emphasis; in scoring, controls per-vintage score aggregation. Available in both fit() and score().