Glosarium AI
Kamus lengkap Kecerdasan Buatan
Forward Chaining
Validation technique where each training set includes all previous observations and the test set contains the immediate future data. Simulates real-time prediction using only information available up to each time point.
Rolling Window Cross-Validation
Method using a fixed-size window that slides over time to create different training and test sets. Maintains a constant sample size while preserving the temporal structure of the data.
Walk-Forward Validation
Iterative validation approach where the model is trained on historical data and then tested on the following period before moving forward in time. Faithfully reproduces real-world conditions for predictive model deployment.
Expanding Window Cross-Validation
Technique where the training window gradually expands to include all available historical data. Particularly suitable when the amount of data increases and all past observations remain relevant.
Time Series Split
Data splitting strategy into training and test sets that strictly respects the time sequence. Implemented in scikit-learn as TimeSeriesSplit to evaluate time series models without information leakage.
Temporal Nested Cross-Validation
Double cross-validation where external validation evaluates model performance while internal validation optimizes hyperparameters. Prevents overfitting while respecting temporal constraints at all levels.
Blocked Cross-Validation
Method introducing separation blocks between training and test sets to reduce temporal correlation. Prevents contamination of temporally close data that could bias performance evaluation.
Time-Shifted Cross-Validation
Approach inserting a time period between training and test sets to simulate a realistic prediction delay. Particularly useful when decisions based on predictions require a certain execution time.
Hierarchical Temporal Cross-Validation
Extension of temporal cross-validation to data with hierarchical structure such as multi-level time series. Simultaneously respects temporal constraints and hierarchical dependencies between data.
Multiple Time Series Cross-Validation
Technique adapted to datasets containing multiple independent time series sharing similar characteristics. Allows generalization across different series while preserving temporal order in each.
Purged Cross-Validation
Method eliminating temporally close observations from training and test sets to reduce dependency. Essential in finance where autocorrelation and market microstructure can contaminate results.
Embargo Cross-Validation
Variant of purged cross-validation extending the exclusion period after each test point. Creates a temporal buffer zone ensuring complete independence between training and testing phases.
Adaptive Temporal Cross-Validation
Dynamic approach adjusting the size of training and test windows according to data characteristics or model performance. Optimizes the use of available data while adapting to temporal changes.
Sliding Window Cross-Validation
Specific implementation of rolling window where the window moves by a fixed step at each iteration. Allows precise control of overlap between successive training and test sets.
Temporal Sampling Cross-Validation
Method periodically sampling test points in time rather than using continuous blocks. Reduces computational cost while maintaining fair representation of different time periods.