🤖 machine learning

Bias-Corrected Adaptive Conformal Inference for Multi-Horizon Time Series Forecasting

This paper introduces Bias-Corrected Adaptive Conformal Inference (BC-ACI), a method that enhances standard ACI for time series forecasting by integrating an online bias estimation mechanism to re-center prediction intervals, thereby significantly reducing interval width under distribution shifts while maintaining rigorous coverage guarantees.

Original authors: Ankit Lade, Sai Krishna J., Indar Kumar

Published 2026-04-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ankit Lade, Sai Krishna J., Indar Kumar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a weather forecaster. Every morning, you predict the temperature for tomorrow. To be helpful, you don't just give a single number (e.g., "It will be 70°F"); you give a range to be safe (e.g., "It will be between 65°F and 75°F"). This range is your prediction interval.

For a long time, statisticians have had a clever tool called Adaptive Conformal Inference (ACI). Think of ACI as a smart, self-adjusting safety net. If your predictions start missing the mark more often than expected, ACI automatically widens your safety net (the interval) to catch the true temperature again. It's great because it guarantees that, over the long run, your net will catch the truth 90% of the time, even if the weather patterns change.

But there's a catch.

The Problem: The "Stubborn Net"

Imagine the weather suddenly changes. Maybe a cold front moves in, and the actual temperature drops by 10 degrees. However, your forecasting model is "stubborn"—it was trained on old data and hasn't been updated. It keeps predicting 70°F, but the real temperature is now 60°F.

Your model has a bias (a consistent error).

Standard ACI's reaction: It sees that it's missing the target. To fix this, it simply widens the net. It might say, "Okay, the temperature is between 50°F and 90°F."
The result: The net is huge! It catches the temperature, yes, but it's so wide it's not very useful. It's like using a fishing net that is 100 feet wide just to catch one small fish. It works, but it's inefficient.

The paper calls this the "symmetric widening problem." ACI can make the net bigger, but it can't move the net to where the fish actually are.

The Solution: BC-ACI (The "Smart Net Mover")

The authors propose a new method called BC-ACI (Bias-Corrected Adaptive Conformal Inference).

Think of BC-ACI as a smart assistant standing next to your fishing net.

It watches the water: It notices that every time you cast the net, the fish are consistently 10 feet to the left of where you aimed.
It moves the net: Instead of just making the net wider, BC-ACI physically shifts the center of the net 10 feet to the left.
The result: Now, you can use a much smaller, tighter net that still catches the fish perfectly.

How It Works (The "Dead Zone" Trick)

You might ask, "What if the assistant gets confused? What if the fish are just swimming randomly and not actually biased?"

If the assistant tries to move the net every time a fish wiggles, the net will shake around uselessly. To prevent this, BC-ACI has a "Dead Zone" (a safety filter).

The Rule: The assistant only moves the net if the bias is big and consistent.
The Metaphor: Imagine a bouncer at a club. If someone is just slightly off-center (noise), the bouncer ignores them. But if someone is clearly, consistently leaning to one side (real bias), the bouncer steps in and moves them.
This ensures that if your model is already perfect (no bias), BC-ACI does nothing and acts exactly like the old ACI. It doesn't break anything that isn't broken.

The Results: When Does It Help?

The paper tested this on four different scenarios:

The "Stubborn" Model: If you use a model that doesn't update itself (like a simple linear regression) and the world changes, BC-ACI is a superhero. It shrinks the prediction intervals by up to 32% while keeping the accuracy the same. It's like getting a tighter, more useful forecast for free.
The "Self-Correcting" Model: If you use a model that fixes its own errors automatically (like a Random Walk), BC-ACI does nothing. It sees there is no persistent bias, so it stays out of the way.
Stable Weather: If nothing changes, BC-ACI is invisible. It adds almost no extra work and doesn't make the intervals worse.

The Big Picture

In the real world, many companies train AI models once and leave them running for months or years without retraining. When the economy shifts, or a new policy changes, or a sensor drifts, these models get "biased."

Old Way (ACI): "We missed the target! Let's make the safety zone huge!" (Inefficient, wide intervals).
New Way (BC-ACI): "We missed the target because we are off-center. Let's move the center and keep the safety zone tight." (Efficient, precise intervals).

In short: BC-ACI is a lightweight upgrade that teaches your prediction intervals to move instead of just stretch. It makes forecasts tighter and more useful when your model gets stuck in a rut, but it promises to do no harm when your model is already doing a good job.

1. Problem Statement

Context: Adaptive Conformal Inference (ACI) is a state-of-the-art method for generating distribution-free prediction intervals for time series under distribution shifts. It guarantees asymptotic coverage by adaptively updating a quantile threshold ( $\alpha_t$ ) based on past miscoverage rates.

The Limitation: Standard ACI constructs prediction intervals that are symmetric around the point prediction ( $\hat{y}_t$ ). It only adjusts the width of the interval by increasing the quantile threshold to maintain coverage.

The Issue: When a deployed forecasting model (e.g., a ridge regression model not retrained for months) encounters a distribution shift (specifically a level shift), it develops persistent prediction bias ( $b \neq 0$ ).
The Consequence: To cover the shifted residuals, standard ACI must symmetrically widen the interval to encompass the bias. This results in a structural inefficiency where the interval width increases by approximately $2|b|$ .
Motivation: An "oracle" that knew the bias could simply shift the interval center by $b$ , maintaining the same coverage with a much narrower interval. The paper aims to approximate this oracle behavior online without retraining the base model.

2. Methodology: BC-ACI

The authors propose BC-ACI (Bias-Corrected ACI), a lightweight extension that estimates and corrects for bias before computing conformal scores.

Core Components:

Per-Horizon Online Bias Estimation:
- For each forecast horizon $h$ , BC-ACI maintains an Exponentially Weighted Moving Average (EWM) of the signed residuals ( $e_t = y_t - \hat{y}_t$ ).
- Update rule: $\hat{b}_{t+1} = (1-\lambda)\hat{b}_t + \lambda e_t$ .
- This allows the system to track slow-moving bias changes specific to different horizons.
Bias-Corrected Nonconformity Scores:
- Before calculating the quantile threshold, the method corrects the residuals: $\tilde{e}_i = e_i - \hat{b}_t$ .
- The interval center is shifted: $\tilde{y}_t = \hat{y}_t + \hat{b}_t$ .
- The prediction interval is constructed as $[\tilde{y}_t - \tilde{q}_t, \tilde{y}_t + \tilde{q}_t]$ , where $\tilde{q}_t$ is the quantile of the corrected absolute residuals $|\tilde{e}_i|$ .
Adaptive Dead-Zone Threshold:
- Problem: On unbiased data, the EWM estimate $\hat{b}_t$ fluctuates around zero due to noise. Applying correction based on this noise degrades interval efficiency.
- Solution: Correction is only applied if the estimated bias exceeds a dynamic threshold: $|\hat{b}_t| > k \cdot \text{MAD}(B)$ , where MAD is the Median Absolute Deviation of the current calibration buffer and $k$ is a user parameter (default 0.5).
- Benefit: This prevents estimation noise from triggering false corrections on stable data, ensuring the method remains neutral when bias is absent.
Multi-Horizon Architecture:
- The method maintains independent buffers, adaptive levels ( $\alpha_t$ ), and bias estimates ( $\hat{b}_t$ ) for each forecast horizon $h$ , acknowledging that bias magnitude and dynamics vary across horizons.

3. Key Contributions & Theoretical Guarantees

Identification of Structural Limitation: The paper proves (Theorem 2) that threshold-only adaptive methods suffer a width overhead of $\Theta(|b|)$ under persistent bias.
Coverage Guarantee: The authors prove (Theorem 3) that BC-ACI preserves the asymptotic marginal coverage guarantee of standard ACI, provided the bias estimate is predictable (measurable with respect to past information).
Efficiency Gains: The method quantifies the width reduction (Theorem 4), showing that under persistent bias, BC-ACI achieves the oracle width $2q_\alpha(0)$ rather than the inflated $2q_\alpha(b)$ .
Graceful Recovery: Theorem 5 demonstrates that on unbiased data, the dead-zone mechanism ensures BC-ACI behaves almost identically to standard ACI (with $<0.2\%$ overhead), preventing performance degradation.
Characterization of Applicability: The paper explicitly defines when the method helps:
- Helps: Models with persistent bias (e.g., Ridge Regression, Neural Networks) after level shifts.
- Neutral: Self-correcting models (e.g., ARIMA, Random Walks) where differencing eliminates bias immediately, or on stationary data.

4. Experimental Results

The authors evaluated BC-ACI across 688 runs involving:

Datasets: 4 synthetic scenarios (Stable, Mean Shift, Volatility Shift, Compound Shift) and 3 real-world datasets (Electricity, Weather, ETTh1).
Models: Ridge Regression (prone to persistent bias) and ARIMA (self-correcting).
Horizons: $h \in \{1, 5, 12, 24\}$ .

Key Findings:

Significant Improvement under Bias: On synthetic data with mean/compound shifts using Ridge regression, BC-ACI reduced the Winkler score by up to 32% compared to standard ACI while maintaining the target 90% coverage.
- Example: On compound shift, Winkler score dropped from 10.57 (ACI) to 7.17 (BC-ACI).
No Harm on Stable Data: On stable control scenarios and real-world datasets (which lacked clear distribution shifts in the evaluation window), BC-ACI showed negligible degradation (Winkler ratio $\approx 1.002$ ), confirming the effectiveness of the dead-zone.
Model Dependency: The improvement was strictly tied to the presence of persistent bias. ARIMA models (which self-correct) showed no significant difference between ACI and BC-ACI, validating the "conditional improvement" hypothesis.
Real Data Nuance: On real-world datasets, BC-ACI showed a slight, statistically insignificant increase in Winkler score (5–15%), attributed to the lack of true distribution shifts in the test windows, causing the EWM to track noise.

5. Significance and Conclusion

Significance:

Practical MLOps: The method addresses a critical real-world scenario where models are deployed for long periods without retraining. In such settings, distribution shifts cause persistent bias that standard conformal methods handle inefficiently.
Efficiency: By correcting the center of the interval rather than just widening it, BC-ACI provides tighter, more actionable prediction intervals without sacrificing coverage guarantees.
Safety: The adaptive dead-zone ensures the method does not degrade performance on stationary data or self-correcting models, making it a safe "plug-and-play" enhancement.

Limitations & Future Work:

The method only corrects location bias, not variance shifts (though ACI handles variance shifts via threshold adaptation).
It relies on a "cold start" period ( $n_0=50$ ) to initialize bias estimates.
Future work should focus on testing with real-world datasets containing known, sharp distribution shifts (e.g., pandemic-era demand) and extending the dead-zone to a formal hypothesis test.

In summary, BC-ACI is a robust, theoretically grounded extension to Adaptive Conformal Inference that specifically targets the inefficiency caused by persistent prediction bias, offering substantial gains in interval tightness for non-retrained models while maintaining strict coverage guarantees.