Smooth Theil-SenI wanted to build a Theil-Sen estimator that could run on more than one bar and produce smoother output than the standard implementation. Theil-Sen regression is a non-parametric method that calculates the median slope between all pairs of points in your dataset, which makes it extremely robust to outliers. The problem is that median operations produce discrete jumps, especially when you're working with limited sample sizes. Every time the median shifts from one value to another, you get a step change in your regression line, which creates visual choppiness that can be distracting even though the underlying calculations are sound.
The solution I ended up going with was convolving a Gaussian kernel around the center of the sorted lists to get a more continuous median estimate. Instead of just picking the middle value or averaging the two middle values when you have an even sample size, the Gaussian kernel weights the values near the center more heavily and smoothly tapers off as you move away from the median position. This creates a weighted average that behaves like a median in terms of robustness but produces much smoother transitions as new data points arrive and the sorted list shifts.
There are variance tradeoffs with this approach since you're no longer using the pure median, but they're minimal in practice. The kernel weighting stays concentrated enough around the center that you retain most of the outlier resistance that makes Theil-Sen useful in the first place. What you gain is a regression line that updates smoothly instead of jumping discretely, which makes it easier to spot genuine trend changes versus just the statistical noise of median recalculation. The smoothness is particularly noticeable when you're running the estimator over longer lookback periods where the sorted list is large enough that small kernel adjustments have less impact on the overall center of mass.
The Gaussian kernel itself is a bell curve centered on the median position, with a standard deviation you can tune to control how much smoothing you want. Tighter kernels stay closer to the pure median behavior and give you more discrete steps. Wider kernels spread the weighting further from the center and produce smoother output at the cost of slightly reduced outlier resistance. The default settings strike a balance that keeps the estimator robust while removing most of the visual jitter.
Running Theil-Sen on multiple bars means calculating slopes between all pairs of points across your lookback window, sorting those slopes, and then applying the Gaussian kernel to find the weighted center of that sorted distribution. This is computationally more expensive than simple moving averages or even standard linear regression, but Pine Script handles it well enough for reasonable lookback lengths. The benefit is that you get a trend estimate that doesn't get thrown off by individual spikes or anomalies in your price data, which is valuable when working with noisy instruments or during volatile periods where traditional regression lines can swing wildly.
The implementation maintains sorted arrays for both the slope calculations and the final kernel weighting, which keeps everything organized and makes the Gaussian convolution straightforward. The kernel weights are precalculated based on the distance from the center position, then applied as multipliers to the sorted slope values before summing to get the final smoothed median slope. That slope gets combined with an intercept calculation to produce the regression line values you see plotted on the chart.
What this really demonstrates is that you can take classical statistical methods like Theil-Sen and adapt them with signal processing techniques like kernel convolution to get behavior that's more suited to real-time visualization. The pure mathematical definition of a median is discrete by nature, but financial charts benefit from smooth, continuous lines that make it easier to track changes over time. By introducing the Gaussian kernel weighting, you preserve the core robustness of the median-based approach while gaining the visual smoothness of methods that use weighted averages. Whether that smoothness is worth the minor variance tradeoff depends on your use case, but for most charting applications, the improved readability makes it a good compromise.
