I Invented 2021 Candle Types to Find the One Holy Grail That Improves Strategies by +1.2 Sharpe
Renko, Heikin Ashi, Point-and-Figure... They have nothing on this brand new candle type.
We spend our trading lives looking at charts, plugging in different indicators, not to realize that we are painting on top of the same old picture. Instead of trying a different color of paint, we need to change the canvas.
And in doing just that, you may have won yourself a spot in the lead.
Here’s the thing:
Everyone is looking at the same charts. Everyone has access to the same information and indicators.
I developed and tested thousands of new candlestick types and found one Holy Grail. And I’m here to show you how I did it and what that new candlestick chart is.
What’s In a Candle
We’re all familiar with the language of the charts: candlesticks.
These things were invented back in the 20s when ticker tape and newspapers were still around. People without access to the tape would just pick up a paper in the morning or after work to see how a stock did over the course of the day.
The analysts chose open, high, low, and closing prices as a quick summary at a glance so you could understand what happened in a day.
Why They Fail
These candlesticks were never supposed to be anything more than daily snapshots. However, just because they were around when computing started to take off, people used them to run data analysis.
Since then, we’ve been kind of stuck with them.
The major problem with candles is that they were never meant to be used for intraday trading. Daily candles start when the market opens and ends when it closes. That barrier makes sense. Arbitrary barriers like 1 hour or 15 minutes lose information. Yet, we still use them because it’s what we have.
Volume Bars
Some progress has been made, like volume bars in the 70s developed by Mandelbrot and Taylor. These bars look at the market in terms of amounts traded within each fixed amount of volume (say 100 shares). One candle = 100 shares traded.
They better represent actual market behavior, so they work better intraday. But they still had problems when it came to things like splits, stocks rising in price and so their share volume goes down, etc.
Renko and Heikin Ashi
Retail brokers popularized some false hope with Renko and Heikin Ashi bars which are a type of technical indicator.
They still use time bars, but somehow look amazing. While many traders fall for them, they are just like moving averages: they look great in hindsight, but until the candle closes, there’s no indication that a trend will go one way or another.
Machine Learning → Dollar Bars
In the age of machine learning, one more bar type was invented: the dollar bar. Just like the volume bar, it measures time as amount of value traded. But instead of counting shares, it counts dollar value traded. This normalizes the volume and fixes the problems arising from splits and other intricacies that original volume bars had.
Above is a good example at how dollar bars can help normalize volume bars in market irregularities.
But that’s it.
No more progress has been done. Instead we get 99.99% of the focus on what lays on top of the bars rather than the bars themselves.
And that got me thinking… is there a better bar out there?
Becoming a Chandler (Candle Maker)
To tackle this, I went ahead and designed a schema for a family of bar types and then the actual bar types themselves. I then generated 60 families ⨉ 60 children using an LLM for a total of 3600 potential solutions.
By using high concurrency and openai/gpt-oss-120b:nitro on OpenRouter, I achieved over 8,000 tokens per second and about 150 solutions per second.
I spent about 10M tokens total which amounted to maybe $2 total. Not bad.
Evaluating The Beasts
Traditional academia uses a Jarque-Bera test which scores how close to normal a data set is distributed. But we don’t care about that. We want profitable bar types. So, I created 4 ‘dummy’ trading strategies:
Bollinger Reversion, mean reversion
Donchain Channel, trend following
SMA Cross, trend following
RSI Reversion, mean reversion
These four strategies were applied to 8 intraday FX assets: AUDUSD, EURUSD, GBPUSD, USDCAD, USDCHF, USDCHN, USDJPY, USDMXN with history between 2020 and 2025.
This gave 4 ⨉ 8 equity curves per bar type, which was more than enough to get a decent read on how good a particular type was compared to another.
To begin, I calculated the new bars on each dataset. A bar type implementation took the 1-minute price data as OHLCV bars as input and was told to resample it however it wanted to. I used 1-minute time bars as the smallest unit of time for simplicity.
Normalizing Bar Frequencies
Each type chooses to resample however it wants. The time that each bar consumes is variable. For a volume based bar, it might take 60 minutes to trade 10,000 shares and then 3 hours for the next 10,000 shares.
So to properly compare this against time bars to see if the type is better at trading signals, I resampled the time bars to fit within the same granularity.
If a type computed 100,000 bars from 2020 → 2015, then I would find what interval I need to set my time bars to so that I got 100,000 time bars. You can see below how that looks on a particular slice of EURUSD price data.
Sharpe Improvement Scoring
Each type was scored on the Sharpe ratio lift that it gave over it’s time bar equivalent strategy. Out of 3600 strategies, 1579 returned duplicate solutions. So 2021 final solutions were scored.
It’s hard to see, but there is a small uplift off to the left. A lot of noise and garbage however was produced. Even so, 179 survive making money and beating their time bar equivalent test.
The space in blue is where our winners lie along with the ‘Pareto front’ which is a bit of a misnomer here but I digress.
Compared on Strategies
Mean Reversion Strategies
Trend Following
Immediately, it was clear to me that the top three winners have some sort of mean reversion capture going on, as they have clean equity curve lines up on the mean reversion strategies and then do consistently badly on the trend following ones.
So I had to ask:
The winners look like mean-reversion instruments. Are they, and is it worth anything?
Digging Deeper
I looked more into the leaderboard and noticed something very important: the edge for a lot of the winners correlated with how ‘fast’ the bars were. For example, bars comprising of on average a small amount of minutes (3 - 10) had a much higher edge on average than ones that were comprised of more minute bars (10 - 100+).
Unsurprisingly, these ‘fast’ bars lose money when fees are added. That’s because they trade too often.
But, we can see that we are capturing something interesting. Our invented clocks (bar types) consistently score better on the reversion strategies than the trend strategies, even if the amount of profit decays with the size of the bar.
This goes against null clocks that sample bars at random, which means that there is something very exploitable here.
The test was expanded to more statistical tests:
Which shows that fast bars breakdown on pretty much every metric besides the xcorr_signal. They have no long term memory via the Hurst exponent, no variance, incorrect autocorrelation, fail the runs test, and fail to close at extremes. These are all effects of the bars just simply being too short in time to capture any interesting information that it can exploit.
Against More Strategies
Furthermore, we hit profitability only after a certain lower threshold of turnover is hit, which makes our case even better. We don’t need fast trading bars. Slow and rich bars contain enough information to be profitable and don’t trade enough to rack up fees.
Further evidence reinforces this as the number of different strategies used to gather data was increased. Reversion rules win and have low turnover. Exactly what we want to see.
The Final Winners
The top three bar types that we originally found in the first half of the study turned out to be the ones selected as well once we dove deeper, which is great for us.
In hold out tests, things get a little iffy for the #2 and #3. But our number one still pulls forward.
Where You Come In
Now that you understand that maybe it’s not just about the indicator, but also the clock, which is a bit counterintuitive. And now maybe you also believe that looking outside of the box is really the way to go when it comes to quant.
That puts you ahead of 99% of quants out there who are looking at the same charts and using the same boring ideas that already have billions of dollars deployed against them.
Be different. Be new.
Getting the Code and Bar Type Formulas
All of this code is available on the private Google Drive so you can go and take the three winning bar type strategies and use them in your own systems and analyses.
Don’t have access? No worries. It’s just a click away. Your subscription supports me and get you code to every blog post I’ve written thus far. Not a bad deal!
Getting All 2021 Systems, Generation Code, etc…
If you want access to all the code that I wrote to recreate this blog post, which includes every single bar implementation, the evaluation scripts, LLM mass generation scripts with the schema and context files used, etc., then you need to be a Professional Member. Adjust your subscription accordingly.
Other massive drops like my previous post on generating entire trading strategies en masse using LLMs are also in the Founders / Professional member’s folder and you can expect other drops to go in there for posts that generate far too much content to contain in a single recreation notebook.
Adjust your subscription, and gain access today.
Happy trading.
Stuart
Well, that didn't go as planned. Anyone else run into issues with CC getting declined when subscribing on Substack?