PROTECTED SOURCE SCRIPT
Neon Juliet - Preview
Actualizado
There is no TLDR, but there is a summary at the end. I strongly encourage to read full description before trying it out. Enjoy!
Background
=========
Having successful and adamant trading systems typically consists of two (oversimplified) elements: signals and risk management system. In most zero-sum games, such as trading, signals must offer an advantage against the market, and risk management system provides a safety mechanism to allow the system to exist in the future. Let me explain.
Say, I have a solid risk management system: it is diversified, with take profit and stop loss thresholds set for low risk, on average I trade less than 3% of my assets, and there’s a loss recovery mechanism, etc. Hypothetically, it’s pristine. Now, let’s trade this portfolio against a flip of a coin, essentially a signal that provides 50% probability of things turning out in my favour. How profitable is such system? My answer: it isn’t. I might be able to sustain this system for some time, but eventually this system is going to have to loosen risk restrictions to stay ahead of the commissions and borrowing costs, resulting in overtime detrimental trend.
Conversely, if the signals provide greater than 50% confidence of things turning out in my favour, but risk management is poor, I’d expect such system to end up in a disaster soon, perhaps after a few euphoric gains. (I’d isolate a top-notch signals, say >90% confidence, in another bucket, but this idealistic system is non-achievable in my practice, so I’ll leave it be)
Neon Juliet was developed to offer an advantage against given markets. Probabilities generated by this model are statistical historical outcomes. This model developed using only price action and is unable to consume any other data or price data across instruments. In other words, it doesn’t know anything you don’t see already on a chart.
Neon J performs best on complex instruments where there’s great diversity of actors and considerable daily volume .
Methodology
==========
In principle, Neon J is based on Bayes’ Theorem. Simply put, prior knowledge of price action ( aka patterns) provides basis for probability of future price action development (ex. long or short trend).
The training process is implemented outside of this script mainly due to Pine Script limitations. This script, however, contains inference portion of the model.
As input for training, daily candle data is used. From this data, feature engineering step of the training develops features, like price average divergence/convergence (think MACD ), price strength (think RSI , ADX ); multiple periods used to diversify long and short patterns. This is done to develop a “state” that is reflective of recent price development. Ex. what we’d call a trend is just a strong and consistent upward price action, but we’d need to look at most recent N candles and their pattern to know that.
Once features are developed, I train a model using Reinforcement Learning technique. Simply put, this technique allows an agent to interact with a trading simulator and take actions (ex. go long, go short, etc.). After many iterations, the agent learns conditions (patterns) that lead to positive outcomes and those that lead to negative outcomes. This learning is quantitative, which means there’s a way to tell which probabilities are strong and which are weak. These probabilities are indicated by this script.
Trained Neon J models are instruments-specific. Meaning, that model for DJI is not compatible with SP500 or any other instrument. Experimentally, I proved that such approach over-performs generalizable models (those that are trained on data from multiple instruments)
Neon J currently only support daily time frame. The limitation is purely practical to reduce the development load and model size.
Results
======
Tests show 60%-70% success rate (on average, some instruments are worse than that, some better) of individual signal when threshold is set to 0.3 (roughly equivalent to 65% probability). This is calculated with Pine Script Strategy with the following entry/exit rules:
Entry when individual signal (a dot) is above 0.3 (long) or below -0.3 (short)
Exit when 14-period smooth signal (a column) is above 0.0 (short exit) or below 0.0 (long exit)
No stop loss or take profit levels.
Pyramiding is set to 100 (to allow unrestricted action of all signals)
All trades are closed on last tested bar (to conclude all signals in-flight)
Percent Profitable is what we take as success rate in the context of this assessment. This number represents how many signals were profitable vs all signals actioned.
It is also worth noting that this assessment was performed on a time period previously unseen by the model. Simply put, we only train a model with data up until date X, then we test starting from date X onward. This ensures that the assessment is unbiased by the model already “knowing” the future. In practice, this gives confidence that future (unknown) market dynamics is going to be representative of our test results.
Be aware, the above “strategy” is not my recommended usage of this signal, it is simply an assessment technique that is meant to be as simple and unconstrained as possible.
How to use this script
================
The script calculates a probability. A term probability here is used in a loose form and means “a numeric value in roughly -1 to 1 space that represents the likelyhood of bullish or bearish price action”. Keep in mind that probability values can go over 1.0 or below -1.0. This is due to the fact that these value are normalized to -1/1 space using 95-percentile (this detail is largely unimportant for usability’s sake).
Indications
--------------
Dots (circles) indicate individual probability value on any given bar. Indicated value on a given bar indicates the probability of future price action. High (positive) values indicate high probability of long action in the future. Low (negative) values indicate high probability of short action in the future. You should interpret future as a gradient (a trend developing slowly over time) instead of being isolated to what’s immediately follows (ex. next bar)
Columns (histogram) provided as convenient view of smoothed probabilities of last N bars. This is controlled by the Smoothing parameter and defaults to 14.
Parameters
---------------
Model parameter is the backbone of this script. It is a required parameter and it is unique for each instrument. Example models provided at the end (see below). This parameter is a long 10000+ character representation of a model.
The script has two additional parameters for configuring interpretation: Threshold and Smoothing.
Threshold controls the level at which values change color (ex. above 0.3, turn neon blue, and below -0.3 turn neon purple).
Smoothing parameter provides a way to smooth out individual probabilities into a exponential moving average with the periods provided. This average is indicated using columns on the indicator.
Model expiration
----------------------
Models are valid for 1 month after training. This is done by design to prevent model deterioration. A month is proven to be a maximum period of time to hold model performance steady. After that, deterioration is likely to occur. Optimal time for model lifetime is 10 days (this is what I use for live trading), and of course most optimal (but unpractical for now) is to re-train daily.
Validity indicated with blue-tinted indicator background, while red-tinted background indicates expired period.
Preview
======
This script is released as a public script for anyone to try. My motives for this release are two-fold:
To subject the model to a variety of conditions, including traders with different experiences trading different instruments (subject to specific models offered of course). Essentially, my own testing is not enough to grasp a full breadths of scenarios. I’d like to harden it and understand where it is strong and where it might fall short (pun intended).
Get an idea on how Neon J might be useful when making trading decision. I tried to make the representation of the signals unconstrained and unopinionated, so there’s room to explore and experiment. I found that Neon J can be packaged in a number of different ways.
At this moment the script is closed-source. I might consider open-sourcing this script in future depending on how much feedback I get from this submission and whether it’d be deemed useful to others.
Summary
=======
Neon J is a set of probabilistic models for predicting future price action with ~65% accuracy. It indicates individual signals (circles) for probability of price action in a foreseeable future, while smoothed signals (columns) are provided for a more dynamic view of probable price action. Blue circle - strong long probability; Purple circle - strong short probability. Blue column - strong long trend ahead or in-progress; Purple column - strong short trend ahead or in-progress.
To use it, copy models below and provide them an input to “model” parameter when applying to a chart. Models are instrument-specific. Only daily (D) charts should be used.
The script is provided for evaluation purposes.
Models!
======
At last, here are the models (a piece of text you need to input in script parameters for each instrument)
DJI:
SP500:
BTCUSD
USDCAD:
For more models, see a link on bio (description length limitation in this description restricts me to publish more).
Unimportant details
===============
Background
=========
Having successful and adamant trading systems typically consists of two (oversimplified) elements: signals and risk management system. In most zero-sum games, such as trading, signals must offer an advantage against the market, and risk management system provides a safety mechanism to allow the system to exist in the future. Let me explain.
Say, I have a solid risk management system: it is diversified, with take profit and stop loss thresholds set for low risk, on average I trade less than 3% of my assets, and there’s a loss recovery mechanism, etc. Hypothetically, it’s pristine. Now, let’s trade this portfolio against a flip of a coin, essentially a signal that provides 50% probability of things turning out in my favour. How profitable is such system? My answer: it isn’t. I might be able to sustain this system for some time, but eventually this system is going to have to loosen risk restrictions to stay ahead of the commissions and borrowing costs, resulting in overtime detrimental trend.
Conversely, if the signals provide greater than 50% confidence of things turning out in my favour, but risk management is poor, I’d expect such system to end up in a disaster soon, perhaps after a few euphoric gains. (I’d isolate a top-notch signals, say >90% confidence, in another bucket, but this idealistic system is non-achievable in my practice, so I’ll leave it be)
Neon Juliet was developed to offer an advantage against given markets. Probabilities generated by this model are statistical historical outcomes. This model developed using only price action and is unable to consume any other data or price data across instruments. In other words, it doesn’t know anything you don’t see already on a chart.
Neon J performs best on complex instruments where there’s great diversity of actors and considerable daily volume .
Methodology
==========
In principle, Neon J is based on Bayes’ Theorem. Simply put, prior knowledge of price action ( aka patterns) provides basis for probability of future price action development (ex. long or short trend).
The training process is implemented outside of this script mainly due to Pine Script limitations. This script, however, contains inference portion of the model.
As input for training, daily candle data is used. From this data, feature engineering step of the training develops features, like price average divergence/convergence (think MACD ), price strength (think RSI , ADX ); multiple periods used to diversify long and short patterns. This is done to develop a “state” that is reflective of recent price development. Ex. what we’d call a trend is just a strong and consistent upward price action, but we’d need to look at most recent N candles and their pattern to know that.
Once features are developed, I train a model using Reinforcement Learning technique. Simply put, this technique allows an agent to interact with a trading simulator and take actions (ex. go long, go short, etc.). After many iterations, the agent learns conditions (patterns) that lead to positive outcomes and those that lead to negative outcomes. This learning is quantitative, which means there’s a way to tell which probabilities are strong and which are weak. These probabilities are indicated by this script.
Trained Neon J models are instruments-specific. Meaning, that model for DJI is not compatible with SP500 or any other instrument. Experimentally, I proved that such approach over-performs generalizable models (those that are trained on data from multiple instruments)
Neon J currently only support daily time frame. The limitation is purely practical to reduce the development load and model size.
Results
======
Tests show 60%-70% success rate (on average, some instruments are worse than that, some better) of individual signal when threshold is set to 0.3 (roughly equivalent to 65% probability). This is calculated with Pine Script Strategy with the following entry/exit rules:
Entry when individual signal (a dot) is above 0.3 (long) or below -0.3 (short)
Exit when 14-period smooth signal (a column) is above 0.0 (short exit) or below 0.0 (long exit)
No stop loss or take profit levels.
Pyramiding is set to 100 (to allow unrestricted action of all signals)
All trades are closed on last tested bar (to conclude all signals in-flight)
Percent Profitable is what we take as success rate in the context of this assessment. This number represents how many signals were profitable vs all signals actioned.
It is also worth noting that this assessment was performed on a time period previously unseen by the model. Simply put, we only train a model with data up until date X, then we test starting from date X onward. This ensures that the assessment is unbiased by the model already “knowing” the future. In practice, this gives confidence that future (unknown) market dynamics is going to be representative of our test results.
Be aware, the above “strategy” is not my recommended usage of this signal, it is simply an assessment technique that is meant to be as simple and unconstrained as possible.
How to use this script
================
The script calculates a probability. A term probability here is used in a loose form and means “a numeric value in roughly -1 to 1 space that represents the likelyhood of bullish or bearish price action”. Keep in mind that probability values can go over 1.0 or below -1.0. This is due to the fact that these value are normalized to -1/1 space using 95-percentile (this detail is largely unimportant for usability’s sake).
Indications
--------------
Dots (circles) indicate individual probability value on any given bar. Indicated value on a given bar indicates the probability of future price action. High (positive) values indicate high probability of long action in the future. Low (negative) values indicate high probability of short action in the future. You should interpret future as a gradient (a trend developing slowly over time) instead of being isolated to what’s immediately follows (ex. next bar)
Columns (histogram) provided as convenient view of smoothed probabilities of last N bars. This is controlled by the Smoothing parameter and defaults to 14.
Parameters
---------------
Model parameter is the backbone of this script. It is a required parameter and it is unique for each instrument. Example models provided at the end (see below). This parameter is a long 10000+ character representation of a model.
The script has two additional parameters for configuring interpretation: Threshold and Smoothing.
Threshold controls the level at which values change color (ex. above 0.3, turn neon blue, and below -0.3 turn neon purple).
Smoothing parameter provides a way to smooth out individual probabilities into a exponential moving average with the periods provided. This average is indicated using columns on the indicator.
Model expiration
----------------------
Models are valid for 1 month after training. This is done by design to prevent model deterioration. A month is proven to be a maximum period of time to hold model performance steady. After that, deterioration is likely to occur. Optimal time for model lifetime is 10 days (this is what I use for live trading), and of course most optimal (but unpractical for now) is to re-train daily.
Validity indicated with blue-tinted indicator background, while red-tinted background indicates expired period.
Preview
======
This script is released as a public script for anyone to try. My motives for this release are two-fold:
To subject the model to a variety of conditions, including traders with different experiences trading different instruments (subject to specific models offered of course). Essentially, my own testing is not enough to grasp a full breadths of scenarios. I’d like to harden it and understand where it is strong and where it might fall short (pun intended).
Get an idea on how Neon J might be useful when making trading decision. I tried to make the representation of the signals unconstrained and unopinionated, so there’s room to explore and experiment. I found that Neon J can be packaged in a number of different ways.
At this moment the script is closed-source. I might consider open-sourcing this script in future depending on how much feedback I get from this submission and whether it’d be deemed useful to others.
Summary
=======
Neon J is a set of probabilistic models for predicting future price action with ~65% accuracy. It indicates individual signals (circles) for probability of price action in a foreseeable future, while smoothed signals (columns) are provided for a more dynamic view of probable price action. Blue circle - strong long probability; Purple circle - strong short probability. Blue column - strong long trend ahead or in-progress; Purple column - strong short trend ahead or in-progress.
To use it, copy models below and provide them an input to “model” parameter when applying to a chart. Models are instrument-specific. Only daily (D) charts should be used.
The script is provided for evaluation purposes.
Models!
======
At last, here are the models (a piece of text you need to input in script parameters for each instrument)
DJI:SP500:BTCUSDUSDCAD:For more models, see a link on bio (description length limitation in this description restricts me to publish more).
Unimportant details
===============
- “Neon” is the project code name, “J” is the iteration (versions “A” to “I” all led to a solid “J”)
- Formatting options here make formatting very difficult, so forgive me poor readability.
Notas de prensa
- Use double standard deviation rule for normalizing probabilities. Before, we were normalizing using P95 of absolute values. This change brings better normalization into -1 to 1 space. It also better represents probabilites and doesn't suffer from extreme low negative values overthrowing the final probability value.- Visual representation of smoothed bars now changes color only for individual bar probabilities and not when smooth value is above/below the thresholds. This makes it easier to read individual signals.
Script protegido
Este script se publica con código cerrado, pero puede utilizarlo libremente. Márquelo como favorito y podrá usarlo en un gráfico. El código fuente no se puede ver ni modificar.
¿Quiere utilizar este script en un gráfico?
Exención de responsabilidad
La información y las publicaciones que ofrecemos, no implican ni constituyen un asesoramiento financiero, ni de inversión, trading o cualquier otro tipo de consejo o recomendación emitida o respaldada por TradingView. Puede obtener información adicional en las Condiciones de uso.