⚡️ tempdisagg

Temporal Disaggregation Models in Python

High-Frequency Estimation from Low-Frequency Data — Modular · Robust · Ready for Production

tempdisagg is a production-ready Python library for temporal disaggregation of time series — transforming low-frequency data into high-frequency estimates while preserving consistency.

It supports all major classical methods — Chow-Lin, Litterman, Denton, Fernández, Uniform — and provides a clean modular architecture inspired by R's tempdisagg, with modern additions:

• 📈 Regression + autoregressive adjustment
• 📉 Differencing & smoothing interpolators
• 🤖 Ensemble prediction engine
• 🧠 Intelligent padding & interpolation
• 🔧 Post-estimation fix for negative values
• 🔄 Optional retropolarization via regression

🔍 Why Temporal Disaggregation?

Official indicators often come in low frequency (e.g. yearly GDP), while economic analysis or forecasting needs monthly or quarterly granularity. tempdisagg fills this gap using consistent econometric techniques to create granular estimates that respect original aggregates.

📚 Methods Implemented

💾 Installation

pip install tempdisagg

🚀 Quick Example

from tempdisagg import TempDisaggModel
import pandas as pd
import numpy as np

# Sample input data (monthly disaggregation of yearly total)
df = pd.DataFrame({
    "Index": [2020]*12 + [2021]*12,
    "Grain": list(range(1, 13)) * 2,
    "y": [1200] + [np.nan]*11 + [1500] + [np.nan]*11,
    "X": np.linspace(100, 200, 24)
})

# Fit model
model = TempDisaggModel(method="chow-lin-opt", conversion="sum")
model.fit(df)

# Predict high-frequency series
y_hat = model.predict()

# Adjust negatives (if any; OPTIONAL)
y_adj = model.adjust_output()

# Show results
model.summary()
model.plot()

⚡ Example with Real Data

import statsmodels.api as sm
from tempdisagg import TempDisaggModel

# Load macroeconomic dataset (quarterly)
macro = sm.datasets.macrodata.load_pandas().data
macro["Index"] = macro["year"].astype(int)
macro["Grain"] = macro["quarter"].astype(int)
macro["X"] = macro["realcons"]

# Aggregate GDP to annual level
gdp_annual = macro.groupby("Index")["realgdp"].mean().reset_index()
gdp_annual.columns = ["Index", "y"]

# Merge back into full frame
df = macro.merge(gdp_annual, on="Index", how="left")[["Index", "Grain", "y", "X"]]

# Fit model and predict
model = TempDisaggModel(method="chow-lin-opt", conversion="average")
model.fit(df)

# Get high-frequency estimates
y_hat = model.predict(full=False)

# Optional: post-estimation adjustment
y_adj = model.adjust_output(full=False)

# Summary and plot
model.summary()
model.plot()

🤖 Ensemble Prediction

Run all models and let the library find the optimal weighted combination.

model = TempDisaggModel(method="ensemble", conversion="sum")
model.fit(df)

model.summary()
model.plot()

Behind the scenes:

• Each method is fitted separately.
• Error metrics (e.g. MAE) are computed.
• Weights are optimized to minimize global error.
• Final prediction is a weighted average across models.

🚫 Negative Value Adjustment

When disaggregation outputs negatives (due to smoothing or regression noise), tempdisagg can correct them without violating consistency.

model.fit(df)
y_hat = model.adjust_output()

Internally:

• Detects negatives in each group.
• Redistributes values proportionally.
• Ensures aggregate values match original data.

🧠 Retropolarizer: Smart Interpolation

For missing values in the target (y), you can activate the Retropolarizer: a module that imputes via regression, proportions, or exponential smoothing.

from tempdisagg import Retropolarizer
retro = Retropolarizer(df = data, new_col = "new", old_col = "old")
df["y_imputed"] = retro.retropolarize(method='proportion')

Or use it inside any model:

model = TempDisaggModel(
    method="chow-lin",
    use_retropolarizer=True,
    retro_method="linear_regression"
)
model.fit(df)

Available methods:

• 'proportion'
• 'linear_regression'
• 'polynomial_regression'
• 'exponential_smoothing'
• 'mlp_regression'

Note: The Retropolarizer is only used to impute missing values in the y column. It is not intended for interpolating the X (indicator) variable.

📘 Input Format

Your data must be in long format:

Index | Grain | y     | X
------|-------|-------|-----
2020  | 1     | 1000  | 10.1
2020  | 2     | 1000  | 11.3
2020  | 3     | 1000  | 12.5
...   | ...   | ...   | ...

🧩 Modular Design

🧪 Testing & Reliability

• ✅ Full test coverage
• ✅ Input validation & fallbacks
• ✅ Padding & missing data supported
• ✅ Consistency validation C @ y_hat ≈ y_l

🔍 API Overview

📦 Dependencies

tempdisagg relies on the following Python libraries:

• pandas – data manipulation
• numpy – numerical operations
• matplotlib – plotting
• scipy and statsmodels – regression and optimization
• scikit-learn – used in Retropolarizer (e.g., MLP imputation)

These packages are automatically installed with pip install tempdisagg.

📄 Cite this work

This project is accompanied by a scientific publication available on arXiv:

Jaime A. Jaramillo-Vera. (2025). tempdisagg: A Python Library for Temporal Disaggregation of Time Series. arXiv:2503.22054 [econ.EM].
🔗 View on arXiv

If you use this library in your research, please cite the paper to support the development and visibility of the project.

@misc{verajaramillo2025tempdisagg,
  title        = {tempdisagg: A Python Library for Temporal Disaggregation of Time Series},
  author       = {Jaime A. Jaramillo-Vera},
  year         = {2025},
  eprint       = {2503.22054},
  archivePrefix = {arXiv},
  primaryClass = {econ.EM},
  url          = {https://arxiv.org/abs/2503.22054}
}

📚 References

• Dagum & Cholette (2006), Benchmarking, Temporal Distribution, and Reconciliation Methods
• Denton (1971), Adjustment of Monthly or Quarterly Series
• Chow & Lin (1971), Best Linear Unbiased Estimation of Missing Observations
• Fernández (1981), Methodological Note on a Monthly Indicator
• Litterman (1983), A Random Walk, Markov Model for Forecasting
• tempdisagg (R package)

📃 License

MIT License — See LICENSE for details.

Developed and maintained by Jaime Vera-Jaramillo — Contributions are welcome ❤️.