I used AI to refactor a Rust RSI indicator. Here is why its results changed

This is a December 2023 development account of changing the RSI calculation used by NexusTrade’s historical implementation. If you were affected by that change, identify RSI-based portfolios, rerun their original backtest periods, compare entry and exit triggers, and review active rules before relying on old results. This article does not identify the platform release or every affected portfolio.

Consider reading and reacting on Medium to spread my reach!

 

Trusting external libraries can be a double-edged sword. When you come across a library that has garnered 574 stars, 103 forks, and is well-documented and tested, it’s natural to have a high level of trust in its reliability. This is especially true when it’s widely used in the community. However, this trust can be misleading. Can a library, even one this popular and seemingly well-vetted, be completely free of major bugs?

During my refactor, the RSI outputs differed from the original library’s tests. I initially described that as a major bug. The historical code below shows changed initialization and averaging behavior, so the discrepancy needs that context before treating one definition as universally correct.

Go to the original ta-rs library

This article shows how a timestamp-window refactor exposed an RSI discrepancy and why the calculation details matter.

Uncovering the Unexpected: A Different Goal Leads to a Major Discovery

Initially, I had no intention of auditing the technical analysis library I was using, trusting in its functionality. My original purpose for delving into the code was entirely different.

In the original library, you define a fixed window length. This is useful for backtesting and live-trading assuming your data is ingested at equal time intervals.

use ta::indicators::ExponentialMovingAverage;
use ta::Next;

// it can return an error, when an invalid length is passed (e.g. 0)
let mut ema = ExponentialMovingAverage::new(3).unwrap();

assert_eq!(ema.next(2.0), 2.0);
assert_eq!(ema.next(5.0), 3.5);
assert_eq!(ema.next(1.0), 2.25);
assert_eq!(ema.next(6.25), 4.25);

The reality is that expecting evenly distributed data in an algorithmic trading platform is quite a stretch. While this assumption could be forced to fit stock trading scenarios, it became clear that it was not viable when combined with cryptocurrency trading. The cryptocurrency API I was integrating with allowed me to fetch minutely, hourly, and daily cryptocurrency data. And I needed the flexibility to work with dynamically-sized windows.

Thus, I decided to fork the repo and fixed the code to work for my particular use case. For example:

use ta::indicators::ExponentialMovingAverage;
use ta::Next;

let mut ema = ExponentialMovingAverage::new(Duration::seconds(3)).unwrap(); // window size of 3 seconds
let now = Utc::now();

assert_eq!(ema.next((now, 2.0)), 2.0);
assert_eq!(ema.next((now + Duration::seconds(1), 5.0)), 3.5);
assert_eq!(ema.next((now + Duration::seconds(2), 1.0)), 2.25);
assert_eq!(ema.next((now + Duration::seconds(3), 6.25)), 4.25);

This dated refactor lets configurations express duration-based windows, such as a three-day SMA or a 30-minute rate of change. It does not establish interval-invariant EMA accuracy: the pinned fork applies a fixed coefficient, 2 divided by (duration in seconds plus 1), once per observation. Removing old timestamps from the stored window does not reweight the current EMA by elapsed time.

Go to my updated ta-rs library

Unearthing the Bug

To reduce the scope, I deleted all technical indicators that weren’t currently implemented in the NexusTrade app. To do the refactor, I used my next-generation AI Configuration Platform, NexusGenAI, to streamline the process.

NexusGenAI allows you to create customized AI workflows. The workflow I created was called “TA-Rust-Refactor” and it contained exactly one prompt.

The “Refactor Rust” prompt used in the Application

While I probably could’ve created a script using NexusGenAI in the backend, I thought it would be faster to simply copy paste the indicators and ask the AI to regenerate the files.

The issue became apparent when I reached the final technical indicator in the library, the Relative Strength Index (RSI). The RSI is a momentum indicator commonly used in technical analysis to assess the magnitude of recent price changes. It helps in determining overbought or oversold conditions in the price of a stock or other asset. Typically, it is presented as an oscillator ranging from zero to 100, indicating potential overbought conditions when above 70 and oversold conditions when below 30.