I built the models.
I ran the data. I wrote the paper.
Comparative Analysis of Regression-Based Forecasting Models for State-Level E-Waste Generation Prediction in India
Abstract
Electrical and electronic waste constitutes one of the most rapidly growing waste streams in India — yet state-level forecasting of e-waste generation has received very little attention. I apply two independent regression models — ordinary least-squares linear regression and second-degree polynomial regression — to CPCB state-level data from 2016 to 2023, then validate predictions against real 2024 values for 13 Indian states. The core finding is that forecasting accuracy depends more on the consistency of a state's reporting record than on which model is used. States with stable historical data (Tamil Nadu, Maharashtra, Uttarakhand) achieved prediction errors as low as 2.37%. States with erratic or imputed data produced errors exceeding 1,000%. The research demonstrates that model complexity is the wrong lever to pull when trying to improve e-waste forecasting — data quality is.
Keywords: e-waste forecasting · regression analysis · India · state-level waste management · CPCB · validation study · data qualityIntroduction
E-waste — the category of discarded electronics including computers, mobile phones, televisions, and household gadgets — is among the fastest-growing waste streams globally. India is one of the world's largest producers of e-waste, and the problem is accelerating as consumer electronics penetration deepens across the country.
The Indian government has responded with regulatory frameworks including the E-Waste (Management) Rules 2022, and the Central Pollution Control Board (CPCB) publishes annual state-level data on e-waste collection and processing. Yet despite the policy significance of these numbers, relatively little work has been done to forecast state-level e-waste generation and then actually test those forecasts against real observed data.
This is the gap this research addresses. I apply two regression specifications — linear and polynomial — to CPCB data from 2016 to 2023, hold out 2024 as a test year (decided before any model was fitted), and compare predicted values against actual 2024 CPCB figures. The research asks five core questions:
- Can regression-based models predict state-level e-waste generation accurately?
- Does linear or polynomial regression perform better — and does the answer vary by state?
- What structural features of a state's historical data predict whether forecasts will be accurate?
- What does this imply for policy investment — better models, or better data collection?
- How do these findings compare with the existing waste forecasting literature?
Methodology
Data: State-level e-waste collection and processing figures from CPCB annual reports, 2016–2024. 13 states with sufficient coverage for modelling. Data up to 2023 used for training; 2024 held out strictly for validation, decided before any fitting.
Data preprocessing: Many CPCB entries have missing values — states where registered recyclers did not file separate annual figures. Missing values were imputed using carry-forward or linear interpolation to preserve the time-series structure, and imputed values were tracked separately in the analysis.
Model 1 — Linear Regression:
Where ŷ is estimated e-waste generation, t is the reporting year, and coefficients are estimated using ordinary least squares.
Model 2 — Polynomial Regression (degree 2):
The squared term captures non-linear growth trajectories. Both models were implemented using scikit-learn 1.5 in Python 3.11.
Performance metrics: Five statistics computed for each model — Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Percentage Error per state, R² (calculated on the out-of-sample 2024 forecast), and Pearson's r. This out-of-sample application of R² and RMSE — rather than the common in-sample fitting — is a methodological distinction from most prior work in this field.
Key Findings
Polynomial wins on MAE; linear wins on RMSE and R²
Polynomial regression achieved lower mean absolute error across the 13 states. Linear regression outperformed on RMSE and R². The divergence is explained by RMSE's sensitivity to single large errors — polynomial regression's accuracy on most states is offset by a handful of extreme outlier predictions.
Excellent accuracy in data-consistent states
Tamil Nadu (2.37% error), Maharashtra, and Uttarakhand all achieved excellent validation. Common factor: stable, consistent historical reporting with low coefficient of variation in their training-period series. The models worked because the data was trustworthy.
Errors exceeding 1,000% in erratic states
Four states produced very high errors (>200%). Two causes: (1) extreme year-to-year variability making any trend line unreliable, and (2) extrapolation error where later data showed sharp discontinuities from earlier trends — caused by regulatory changes or sudden reporting shifts, not forecasting model failure.
Data quality predicts accuracy better than model choice
Tamil Nadu (training CV = 0.34, error = 2.37%) vs Andhra Pradesh (high CV, error >200%). The coefficient of variation of a state's historical series is a strong predictor of whether any model will validate well — more predictive than which regression type is used.
Polynomial better for 9 of 13 states
Polynomial regression was the better-performing model for Tamil Nadu, Maharashtra, Punjab, Telangana, Kerala, Rajasthan, Karnataka, Chhattisgarh, and Andhra Pradesh. Linear regression performed better for the remaining four states — no single model dominates across all geographies.
First out-of-sample validation of this kind
Most waste forecasting papers in the literature use in-sample fit statistics (R² on training data). This study uses a strictly held-out 2024 test year — a methodological distinction that makes the accuracy estimates genuinely informative rather than circular.
Conclusion
This research asked whether regression models can forecast state-level e-waste generation in India accurately. The answer is: yes, but only for states where historical reporting is consistent. For states with erratic data, even the most carefully constructed model cannot compensate for the underlying information problem.
The main finding in one sentence
Forecasting accuracy in this domain depends not on the complexity of the model, but on the coherence and consistency of the reporting data that feeds it — and this dependency is measurable in advance using the coefficient of variation of the historical series.
The policy implication is concrete: investment aimed at improving e-waste management forecasting should go first into improving state-level CPCB reporting infrastructure — not into more sophisticated modelling. Only after reporting records become consistent does the choice of model begin to matter.
Limitations of the study include: 13 states rather than all Indian states, only two regression specifications tested, and a single held-out validation year. Future work could extend to more states, apply time-series models (ARIMA, grey models), and incorporate socioeconomic predictors such as GDP per capita and internet penetration as covariates.
The study contributes a replicable validation framework — train on 2016–2023, test on 2024 actuals — that can be reused as new CPCB data becomes available annually, making it a living benchmark for state-level e-waste forecasting in India.
Why this research and Battery Bin connect
Battery Bin collects batteries at the school level. Threco processes them at the national certified level. This research provides the quantitative backdrop to both — understanding at a data-science level exactly how big the e-waste problem is state by state, how it is growing, and where current forecasting methods succeed or fail. The practitioner and the researcher are the same person.
CPCB Data · Python · scikit-learn · State-level India · Out-of-sample validation · 13 states