Skip to content
All library documents

Why Loess Fits Can Mislead in Real-Time Trading

Article Quant Q&A · Author: Nikos
Read the original

Summary

The document asks whether a locally weighted regression, or loess, fit to historical data can support reliable future trading decisions. It offers a caution from a trader who reports testing lowess-based systems over an extended period without finding one that worked. The central issue is that the smoother can use observations on both sides of a point, so its historical curve changes when later bars arrive. A chart built from the full sample can therefore look much better than the signal that would have been available live.

The post gives no specific strategy, dataset, performance statistics, or successful example, and it does not establish that every possible loess-based approach fails. Its practical lesson is to distinguish retrospective smoothing from a causal, point-in-time signal. Any trading evaluation would need to reconstruct what the model could have known at each decision time and account for revisions as new observations arrive.

Key ideas

  • A loess curve may incorporate future observations when smoothing historical points.
  • Its appearance on a complete chart can overstate the quality of a real-time signal.
  • One contributor reports unsuccessful testing but provides no strategy details or performance evidence.
  • Trading use requires evaluating only information available at each decision time.

Tags

StatisticsBacktestingTechnical indicators
Full text
# How to construct a deterministic trading model based on a loess (local regression) model?


# How to construct a deterministic trading model based on a loess (local regression) model?












Given data that has been fit to a loess model, can you make reliable decisions on future trades given a good past fit? Has anyone here done so and can give an example of their use case?

I am yet to decide what data set to use.

## Answer by user3511894 (score 1)

https://quant.stackexchange.com/a/16597

I've been testing trading systems with lowess for a long time and have yet to find something that works. Lowess is a adaptive, so it will change with the addition of future bars ( it is the best fit taking into account past and future data points). If you plot a lowess on a chart, it is a perfect fit...not so much in real time.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.