autotrend 0.2.0

Local Linear Trend Extraction for Time Series

📈 AutoTrend: Local Linear Trend Extraction

AutoTrend is a lightweight, iterative method for extracting local linear trends from time series data. Unlike traditional sliding window approaches that fit a model at every point, AutoTrend achieves computational efficiency by training a single linear regression model per focus region and extending the trend forward, measuring prediction errors without repeated model fitting.

🚀 Demo: Google Colab

---

🚀 Quick Start

import numpy as np
from autotrend import decompose_llt, plot_full_decomposition

# Generate or load your time series
sequence = np.sin(np.linspace(0, 50, 500)) + np.linspace(0, 5, 500)

# Run LLT decomposition
result = decompose_llt(
    seq=sequence,
    max_models=5,
    window_size=10,
    error_percentile=40
)

# Visualize results
plot_full_decomposition(sequence, result)

# Access results
print(f"Number of iterations: {result.get_num_iterations()}")
print(f"Trend segments: {result.get_trend_segments()}")

Output:

• result.trend_marks: Array indicating which iteration labeled each point
• result.prediction_marks: Predicted values for each point
• result.models: List of LinearRegression models from each iteration
• result.process_logs: Detailed logs for visualization

---

💡 Core Concept

The Problem

Traditional sliding window regression methods fit a new model at every time point, leading to high computational costs. Change point detection methods often require complex algorithms and parameter tuning.

The Solution

AutoTrend uses an iterative, focus-based approach:

1. Single Model per Region: Train one linear regression model at the start of each focus region
2. Trend Extension: Extend the trend line forward without retraining
3. Error-Based Refinement: Identify high-error points and focus on them in the next iteration
4. Adaptive Segmentation: Automatically discover trend boundaries based on prediction error

Key Advantages

✅ Computationally Efficient: Minimal model training compared to full sliding windows
✅ Adaptive: Automatically discovers trend boundaries without predefined change points
✅ Interpretable: Clear linear segments with explicit slopes and intercepts
✅ Flexible: Adjustable error thresholds and iteration limits
✅ Lightweight: No complex optimization or parameter search required

---

⚙️ Algorithm Overview

Input

• Sequence: Univariate time series y = [y₀, y₁, ..., yₜ]
• Parameters:
  • window_size: Size of training window (default: 10)
  • max_models: Maximum iterations (default: 5)
  • error_percentile: Error threshold percentile (default: 40)
  • percentile_step: Increment per iteration (default: 0)
  • update_threshold: Whether to update threshold each iteration (default: False)

Process

Step 1: Initialization

Define initial focus targets covering all predictable points:

focus_targets = [window_size, window_size+1, ..., T-1]

Step 2: Train Linear Model

For each iteration, train a model on the first window of the focus region:

X_train = [0, 1, ..., window_size-1]
y_train = sequence[start:end]
model = LinearRegression().fit(X_train, y_train)

Step 3: Extend Trend and Measure Error

Predict forward using the trained model's trend offset:

Δ = ŷ_window_size - y_start
ŷ_t = y_(t-window_size) + Δ
error_t = |y_t - ŷ_t|

Step 4: Segment by Error Threshold

threshold = percentile(errors, error_percentile)
low_error_points = {t | error_t ≤ threshold}
high_error_points = {t | error_t > threshold}

• Low error points: Assigned to current iteration, marked as resolved
• High error points: Become focus targets for next iteration

Step 5: Iterate

Repeat Steps 2-4 on high-error regions until:

• All points meet the error criterion, OR
• Maximum iterations reached

Output

LLTResult(
    trend_marks: np.ndarray,      # Iteration labels for each point
    prediction_marks: np.ndarray,  # Predicted values
    models: List[LinearRegression], # Trained models per iteration
    process_logs: List[Tuple]      # Detailed iteration logs
)

---

📂 Directory Structure

autotrend/
├── autotrend/
│   ├── __init__.py                    # Main package exports
│   ├── core/
│   │   ├── __init__.py
│   │   ├── local_linear_trend.py      # Core LLT algorithm
│   │   └── utility.py                 # Helper functions (extract_ranges, split_by_gap)
│   ├── data/
│   │   ├── __init__.py
│   │   ├── sythn_data/
│   │   │   ├── __init__.py
│   │   │   ├── generate_simple_wave.py          # Stationary sine wave generator
│   │   │   ├── generate_nonstationary_wave.py   # Amplitude-modulated wave generator
│   │   │   └── generate_piecewise_linear.py     # Piecewise linear sequence generator
│   │   └── datasets/                  # Future: Real-world dataset loaders
│   ├── visualization/
│   │   ├── __init__.py
│   │   ├── plot.py                    # Main plotting module
│   │   ├── plot_error.py              # Error analysis visualization
│   │   ├── plot_slope.py              # Slope comparison plots
│   │   ├── plot_full_decomposition.py # Full decomposition view
│   │   ├── plot_iteration_grid.py     # Iteration grid visualization
│   │   └── plot_model_statistics.py   # Model statistics plots
│   └── decomposition/
│       └── __init__.py                # Future: Trend-seasonal decomposition
├── demo/
│   ├── demo_utils.py                  # Demo configuration and utilities
│   ├── simple_wave.py                 # Sine wave demo
│   ├── piecewise_linear.py            # Piecewise linear demo
│   └── run_all.py                     # Run all demos
├── output/                            # Generated plots and logs
│   ├── simple_wave/
│   └── piecewise_linear/
├── README.md
└── requirements.txt