Lorentzian Classification Strategy Based in the model of Machine learning: Lorentzian Classification by @jdehorty, you will be able to get into trending moves and get interesting entries in the market with this strategy. I also put some new features for better backtesting results!
Backtesting context: 2022-07-19 to 2023-04-14 of US500 1H by PEPPERSTONE. Commissions: 0.03% for each entry, 0.03% for each exit. Risk per trade: 2.5% of the total account
For this strategy, 3 indicators are used:
Machine learning: Lorentzian Classification by @jdehorty
One Ema of 200 periods for identifying the trend
Supertrend indicator as a filter for some exits
Atr stop loss from Gatherio
Trade conditions:
For longs:
Close price is above 200 Ema
Lorentzian Classification indicates a buying signal
This gives us our long signal. Stop loss will be determined by atr stop loss (white point), break even(blue point) by a risk/reward ratio of 1:1 and take profit of 3:1 where half position will be closed. This will be showed as buy.
The other half will be closed when the model indicates a selling signal or Supertrend indicator gives a bearish signal. This will be showed as cl buy.
For shorts:
Close price is under 200 Ema
Lorentzian Classification indicates a selling signal
This gives us our short signal. Stop loss will be determined by atr stop loss (white point), break even(blue point) by a risk/reward ratio of 1:1 and take profit of 3:1 where half position will be closed. This will be showed as sell.
The other half will be closed when the model indicates a buying signal or Supertrend indicator gives a bullish signal. This will be showed as cl sell.
Risk management
To calculate the amount of the position you will use just a small percent of your initial capital for the strategy and you will use the atr stop loss or last swing for this.
Example: You have 1000 usd and you just want to risk 2,5% of your account, there is a buy signal at price of 4,000 usd. The stop loss price from atr stop loss or last swing is 3,900. You calculate the distance in percent between 4,000 and 3,900. In this case, that distance would be of 2.50%. Then, you calculate your position by this way: (initial or current capital * risk per trade of your account) / (stop loss distance).
Using these values on the formula: (1000*2,5%)/(2,5%) = 1000usd. It means, you have to use 1000 usd for risking 2.5% of your account.
We will use this risk management for applying compound interest.
> In settings, with position amount calculator, you can enter the amount in usd of your account and the amount in percentage for risking per trade of the account. You will see this value in green color in the upper left corner that shows the amount in usd to use for risking the specific percentage of your account.
> You can also choose a fixed amount, so you will have to activate fixed amount in risk management for trades and set the fixed amount for backtesting.
Script functions
Inside of settings, you will find some utilities for display atr stop loss, break evens, positions, signals, indicators, a table of some stats from backtesting, etc.
You will find the settings for risk management at the end of the script if you want to change something or trying new values for other assets for backtesting.
If you want to change the initial capital for backtest the strategy, go to properties, and also enter the commisions of your exchange and slippage for more realistic results.
In risk managment you can find an option called "Use leverage ?", activate this if you want to backtest using leverage, which means that in case of not having enough money for risking the % determined by you of your account using your initial capital, you will use leverage for using the enough amount for risking that % of your acount in a buy position. Otherwise, the amount will be limited by your initial/current capital
I also added a function for backtesting if you had added or withdrawn money frequently:
Adding money: You can choose how often you want to add money (Monthly, yearly, daily or weekly). Then a fixed amount of money and activate or deactivate this function
Withdraw money: You can choose if you want to withdraw a fixed amount or a percentage of earnings. Then you can choose a fixed amount of money, the period of time and activate or deactivate this function. Also, the percentage of earnings if you choosed this option.
Some other assets where strategy has worked
BTCUSD 4H, 1D
ETHUSD 4H, 1D
BNBUSD 4H
SPX 1D
BANKNIFTY 4H, 15 min
Some things to consider
USE UNDER YOUR OWN RISK. PAST RESULTS DO NOT REPRESENT THE FUTURE.
DEPENDING OF % ACCOUNT RISK PER TRADE, YOU COULD REQUIRE LEVERAGE FOR OPEN SOME POSITIONS, SO PLEASE, BE CAREFULL AND USE CORRECTLY THE RISK MANAGEMENT
Do not forget to change commissions and other parameters related with back testing results!. If you have problems loading the script reduce max bars back number in general settings
Strategies for trending markets use to have more looses than wins and it takes a long time to get profits, so do not forget to be patient and consistent !
Please, visit the post from @jdehorty called Machine Learning: Lorentzian Classification for a better understanding of his script!
Any support and boosts will be well received. If you have any question, do not doubt to ask!
Machinelearning
Machine Learning: Lorentzian Classification█ OVERVIEW
A Lorentzian Distance Classifier (LDC) is a Machine Learning classification algorithm capable of categorizing historical data from a multi-dimensional feature space. This indicator demonstrates how Lorentzian Classification can also be used to predict the direction of future price movements when used as the distance metric for a novel implementation of an Approximate Nearest Neighbors (ANN) algorithm.
█ BACKGROUND
In physics, Lorentzian space is perhaps best known for its role in describing the curvature of space-time in Einstein's theory of General Relativity (2). Interestingly, however, this abstract concept from theoretical physics also has tangible real-world applications in trading.
Recently, it was hypothesized that Lorentzian space was also well-suited for analyzing time-series data (4), (5). This hypothesis has been supported by several empirical studies that demonstrate that Lorentzian distance is more robust to outliers and noise than the more commonly used Euclidean distance (1), (3), (6). Furthermore, Lorentzian distance was also shown to outperform dozens of other highly regarded distance metrics, including Manhattan distance, Bhattacharyya similarity, and Cosine similarity (1), (3). Outside of Dynamic Time Warping based approaches, which are unfortunately too computationally intensive for PineScript at this time, the Lorentzian Distance metric consistently scores the highest mean accuracy over a wide variety of time series data sets (1).
Euclidean distance is commonly used as the default distance metric for NN-based search algorithms, but it may not always be the best choice when dealing with financial market data. This is because financial market data can be significantly impacted by proximity to major world events such as FOMC Meetings and Black Swan events. This event-based distortion of market data can be framed as similar to the gravitational warping caused by a massive object on the space-time continuum. For financial markets, the analogous continuum that experiences warping can be referred to as "price-time".
Below is a side-by-side comparison of how neighborhoods of similar historical points appear in three-dimensional Euclidean Space and Lorentzian Space:
This figure demonstrates how Lorentzian space can better accommodate the warping of price-time since the Lorentzian distance function compresses the Euclidean neighborhood in such a way that the new neighborhood distribution in Lorentzian space tends to cluster around each of the major feature axes in addition to the origin itself. This means that, even though some nearest neighbors will be the same regardless of the distance metric used, Lorentzian space will also allow for the consideration of historical points that would otherwise never be considered with a Euclidean distance metric.
Intuitively, the advantage inherent in the Lorentzian distance metric makes sense. For example, it is logical that the price action that occurs in the hours after Chairman Powell finishes delivering a speech would resemble at least some of the previous times when he finished delivering a speech. This may be true regardless of other factors, such as whether or not the market was overbought or oversold at the time or if the macro conditions were more bullish or bearish overall. These historical reference points are extremely valuable for predictive models, yet the Euclidean distance metric would miss these neighbors entirely, often in favor of irrelevant data points from the day before the event. By using Lorentzian distance as a metric, the ML model is instead able to consider the warping of price-time caused by the event and, ultimately, transcend the temporal bias imposed on it by the time series.
For more information on the implementation details of the Approximate Nearest Neighbors (ANN) algorithm used in this indicator, please refer to the detailed comments in the source code.
█ HOW TO USE
Below is an explanatory breakdown of the different parts of this indicator as it appears in the interface:
Below is an explanation of the different settings for this indicator:
General Settings:
Source - This has a default value of "hlc3" and is used to control the input data source.
Neighbors Count - This has a default value of 8, a minimum value of 1, a maximum value of 100, and a step of 1. It is used to control the number of neighbors to consider.
Max Bars Back - This has a default value of 2000.
Feature Count - This has a default value of 5, a minimum value of 2, and a maximum value of 5. It controls the number of features to use for ML predictions.
Color Compression - This has a default value of 1, a minimum value of 1, and a maximum value of 10. It is used to control the compression factor for adjusting the intensity of the color scale.
Show Exits - This has a default value of false. It controls whether to show the exit threshold on the chart.
Use Dynamic Exits - This has a default value of false. It is used to control whether to attempt to let profits ride by dynamically adjusting the exit threshold based on kernel regression.
Feature Engineering Settings:
Note: The Feature Engineering section is for fine-tuning the features used for ML predictions. The default values are optimized for the 4H to 12H timeframes for most charts, but they should also work reasonably well for other timeframes. By default, the model can support features that accept two parameters (Parameter A and Parameter B, respectively). Even though there are only 4 features provided by default, the same feature with different settings counts as two separate features. If the feature only accepts one parameter, then the second parameter will default to EMA-based smoothing with a default value of 1. These features represent the most effective combination I have encountered in my testing, but additional features may be added as additional options in the future.
Feature 1 - This has a default value of "RSI" and options are: "RSI", "WT", "CCI", "ADX".
Feature 2 - This has a default value of "WT" and options are: "RSI", "WT", "CCI", "ADX".
Feature 3 - This has a default value of "CCI" and options are: "RSI", "WT", "CCI", "ADX".
Feature 4 - This has a default value of "ADX" and options are: "RSI", "WT", "CCI", "ADX".
Feature 5 - This has a default value of "RSI" and options are: "RSI", "WT", "CCI", "ADX".
Filters Settings:
Use Volatility Filter - This has a default value of true. It is used to control whether to use the volatility filter.
Use Regime Filter - This has a default value of true. It is used to control whether to use the trend detection filter.
Use ADX Filter - This has a default value of false. It is used to control whether to use the ADX filter.
Regime Threshold - This has a default value of -0.1, a minimum value of -10, a maximum value of 10, and a step of 0.1. It is used to control the Regime Detection filter for detecting Trending/Ranging markets.
ADX Threshold - This has a default value of 20, a minimum value of 0, a maximum value of 100, and a step of 1. It is used to control the threshold for detecting Trending/Ranging markets.
Kernel Regression Settings:
Trade with Kernel - This has a default value of true. It is used to control whether to trade with the kernel.
Show Kernel Estimate - This has a default value of true. It is used to control whether to show the kernel estimate.
Lookback Window - This has a default value of 8 and a minimum value of 3. It is used to control the number of bars used for the estimation. Recommended range: 3-50
Relative Weighting - This has a default value of 8 and a step size of 0.25. It is used to control the relative weighting of time frames. Recommended range: 0.25-25
Start Regression at Bar - This has a default value of 25. It is used to control the bar index on which to start regression. Recommended range: 0-25
Display Settings:
Show Bar Colors - This has a default value of true. It is used to control whether to show the bar colors.
Show Bar Prediction Values - This has a default value of true. It controls whether to show the ML model's evaluation of each bar as an integer.
Use ATR Offset - This has a default value of false. It controls whether to use the ATR offset instead of the bar prediction offset.
Bar Prediction Offset - This has a default value of 0 and a minimum value of 0. It is used to control the offset of the bar predictions as a percentage from the bar high or close.
Backtesting Settings:
Show Backtest Results - This has a default value of true. It is used to control whether to display the win rate of the given configuration.
█ WORKS CITED
(1) R. Giusti and G. E. A. P. A. Batista, "An Empirical Comparison of Dissimilarity Measures for Time Series Classification," 2013 Brazilian Conference on Intelligent Systems, Oct. 2013, DOI: 10.1109/bracis.2013.22.
(2) Y. Kerimbekov, H. Ş. Bilge, and H. H. Uğurlu, "The use of Lorentzian distance metric in classification problems," Pattern Recognition Letters, vol. 84, 170–176, Dec. 2016, DOI: 10.1016/j.patrec.2016.09.006.
(3) A. Bagnall, A. Bostrom, J. Large, and J. Lines, "The Great Time Series Classification Bake Off: An Experimental Evaluation of Recently Proposed Algorithms." ResearchGate, Feb. 04, 2016.
(4) H. Ş. Bilge, Yerzhan Kerimbekov, and Hasan Hüseyin Uğurlu, "A new classification method by using Lorentzian distance metric," ResearchGate, Sep. 02, 2015.
(5) Y. Kerimbekov and H. Şakir Bilge, "Lorentzian Distance Classifier for Multiple Features," Proceedings of the 6th International Conference on Pattern Recognition Applications and Methods, 2017, DOI: 10.5220/0006197004930501.
(6) V. Surya Prasath et al., "Effects of Distance Measure Choice on KNN Classifier Performance - A Review." .
█ ACKNOWLEDGEMENTS
@veryfid - For many invaluable insights, discussions, and advice that helped to shape this project.
@capissimo - For open sourcing his interesting ideas regarding various KNN implementations in PineScript, several of which helped inspire my original undertaking of this project.
@RikkiTavi - For many invaluable physics-related conversations and for his helping me develop a mechanism for visualizing various distance algorithms in 3D using JavaScript
@jlaurel - For invaluable literature recommendations that helped me to understand the underlying subject matter of this project.
@annutara - For help in beta-testing this indicator and for sharing many helpful ideas and insights early on in its development.
@jasontaylor7 - For helping to beta-test this indicator and for many helpful conversations that helped to shape my backtesting workflow
@meddymarkusvanhala - For helping to beta-test this indicator
@dlbnext - For incredibly detailed backtesting testing of this indicator and for sharing numerous ideas on how the user experience could be improved.
MLExtensionsLibrary "MLExtensions"
normalizeDeriv(src, quadraticMeanLength)
Returns the smoothed hyperbolic tangent of the input series.
Parameters:
src : The input series (i.e., the first-order derivative for price).
quadraticMeanLength : The length of the quadratic mean (RMS).
Returns: nDeriv The normalized derivative of the input series.
normalize(src, min, max)
Rescales a source value with an unbounded range to a target range.
Parameters:
src : The input series
min : The minimum value of the unbounded range
max : The maximum value of the unbounded range
Returns: The normalized series
rescale(src, oldMin, oldMax, newMin, newMax)
Rescales a source value with a bounded range to anther bounded range
Parameters:
src : The input series
oldMin : The minimum value of the range to rescale from
oldMax : The maximum value of the range to rescale from
newMin : The minimum value of the range to rescale to
newMax : The maximum value of the range to rescale to
Returns: The rescaled series
color_green(prediction)
Assigns varying shades of the color green based on the KNN classification
Parameters:
prediction : Value (int|float) of the prediction
Returns: color
color_red(prediction)
Assigns varying shades of the color red based on the KNN classification
Parameters:
prediction : Value of the prediction
Returns: color
tanh(src)
Returns the the hyperbolic tangent of the input series. The sigmoid-like hyperbolic tangent function is used to compress the input to a value between -1 and 1.
Parameters:
src : The input series (i.e., the normalized derivative).
Returns: tanh The hyperbolic tangent of the input series.
dualPoleFilter(src, lookback)
Returns the smoothed hyperbolic tangent of the input series.
Parameters:
src : The input series (i.e., the hyperbolic tangent).
lookback : The lookback window for the smoothing.
Returns: filter The smoothed hyperbolic tangent of the input series.
tanhTransform(src, smoothingFrequency, quadraticMeanLength)
Returns the tanh transform of the input series.
Parameters:
src : The input series (i.e., the result of the tanh calculation).
smoothingFrequency
quadraticMeanLength
Returns: signal The smoothed hyperbolic tangent transform of the input series.
n_rsi(src, n1, n2)
Returns the normalized RSI ideal for use in ML algorithms.
Parameters:
src : The input series (i.e., the result of the RSI calculation).
n1 : The length of the RSI.
n2 : The smoothing length of the RSI.
Returns: signal The normalized RSI.
n_cci(src, n1, n2)
Returns the normalized CCI ideal for use in ML algorithms.
Parameters:
src : The input series (i.e., the result of the CCI calculation).
n1 : The length of the CCI.
n2 : The smoothing length of the CCI.
Returns: signal The normalized CCI.
n_wt(src, n1, n2)
Returns the normalized WaveTrend Classic series ideal for use in ML algorithms.
Parameters:
src : The input series (i.e., the result of the WaveTrend Classic calculation).
n1
n2
Returns: signal The normalized WaveTrend Classic series.
n_adx(highSrc, lowSrc, closeSrc, n1)
Returns the normalized ADX ideal for use in ML algorithms.
Parameters:
highSrc : The input series for the high price.
lowSrc : The input series for the low price.
closeSrc : The input series for the close price.
n1 : The length of the ADX.
regime_filter(src, threshold, useRegimeFilter)
Parameters:
src
threshold
useRegimeFilter
filter_adx(src, length, adxThreshold, useAdxFilter)
filter_adx
Parameters:
src : The source series.
length : The length of the ADX.
adxThreshold : The ADX threshold.
useAdxFilter : Whether to use the ADX filter.
Returns: The ADX.
filter_volatility(minLength, maxLength, useVolatilityFilter)
filter_volatility
Parameters:
minLength : The minimum length of the ATR.
maxLength : The maximum length of the ATR.
useVolatilityFilter : Whether to use the volatility filter.
Returns: Boolean indicating whether or not to let the signal pass through the filter.
backtest(high, low, open, startLongTrade, endLongTrade, startShortTrade, endShortTrade, isStopLossHit, maxBarsBackIndex, thisBarIndex)
Performs a basic backtest using the specified parameters and conditions.
Parameters:
high : The input series for the high price.
low : The input series for the low price.
open : The input series for the open price.
startLongTrade : The series of conditions that indicate the start of a long trade.`
endLongTrade : The series of conditions that indicate the end of a long trade.
startShortTrade : The series of conditions that indicate the start of a short trade.
endShortTrade : The series of conditions that indicate the end of a short trade.
isStopLossHit : The stop loss hit indicator.
maxBarsBackIndex : The maximum number of bars to go back in the backtest.
thisBarIndex : The current bar index.
Returns: A tuple containing backtest values
init_table()
init_table()
Returns: tbl The backtest results.
update_table(tbl, tradeStatsHeader, totalTrades, totalWins, totalLosses, winLossRatio, winrate, stopLosses)
update_table(tbl, tradeStats)
Parameters:
tbl : The backtest results table.
tradeStatsHeader : The trade stats header.
totalTrades : The total number of trades.
totalWins : The total number of wins.
totalLosses : The total number of losses.
winLossRatio : The win loss ratio.
winrate : The winrate.
stopLosses : The total number of stop losses.
Returns: Updated backtest results table.
kNNLibrary "kNN"
Collection of experimental kNN functions. This is a work in progress, an improvement upon my original kNN script:
The script can be recreated with this library. Unlike the original script, that used multiple arrays, this has been reworked with the new Pine Script matrix features.
To make a kNN prediction, the following data should be supplied to the wrapper:
kNN : filter type. Right now either Binary or Percent . Binary works like in the original script: the system stores whether the price has increased (+1) or decreased (-1) since the previous knnStore event (called when either long or short condition is supplied). Percent works the same, but the values stored are the difference of prices in percents. That way larger differences in prices would give higher scores.
k : number k. This is how many nearest neighbors are to be selected (and summed up to get the result).
skew : kNN minimum difference. Normally, the prediction is done with a simple majority of the neighbor votes. If skew is given, then more than a simple majority is needed for a prediction. This also means that there are inputs for which no prediction would be given (if the majority votes are between -skew and +skew). Note that in Percent mode more profitable trades will have higher voting power.
depth : kNN matrix size limit. Originally, the whole available history of trades was used to make a prediction. This not only requires more computational power, but also neglects the fact that the market conditions are changing. This setting restricts the memory matrix to a finite number of past trades.
price : price series
long : long condition. True if the long conditions are met, but filters are not yet applied. For example, in my original script, trades are only made on crossings of fast and slow MAs. So, whenever it is possible to go long, this value is set true. False otherwise.
short : short condition. Same as long , but for short condition.
store : whether the inputs should be stored. Additional filters may be applied to prevent bad trades (for example, trend-based filters), so if you only need to consult kNN without storing the trade, this should be set to false.
feature1 : current value of feature 1. A feature in this case is some kind of data derived from the price. Different features may be used to analyse the price series. For example, oscillator values. Not all of them may be used for kNN prediction. As the current kNN implementation is 2-dimensional, only two features can be used.
feature2 : current value of feature 2.
The wrapper returns a tuple: [ longOK, shortOK ]. This is a pair of filters. When longOK is true, then kNN predicts a long trade may be taken. When shortOK is true, then kNN predicts a short trade may be taken. The kNN filters are returned whenever long or short conditions are met. The trade is supposed to happen when long or short conditions are met and when the kNN filter for the desired direction is true.
Exported functions :
knnStore(knn, p1, p2, src, maxrows)
Store the previous trade; buffer the current one until results are in. Results are binary: up/down
Parameters:
knn : knn matrix
p1 : feature 1 value
p2 : feature 2 value
src : current price
maxrows : limit the matrix size to this number of rows (0 of no limit)
Returns: modified knn matrix
knnStorePercent(knn, p1, p2, src, maxrows)
Store the previous trade; buffer the current one until results are in. Results are in percents
Parameters:
knn : knn matrix
p1 : feature 1 value
p2 : feature 2 value
src : current price
maxrows : limit the matrix size to this number of rows (0 of no limit)
Returns: modified knn matrix
knnGet(distance, result)
Get neighbours by getting k results with the smallest distances
Parameters:
distance : distance array
result : result array
Returns: array slice of k results
knnDistance(knn, p1, p2)
Create a distance array from the two given parameters
Parameters:
knn : knn matrix
p1 : feature 1 value
p2 : feature 2 value
Returns: distance array
knnSum(knn, p1, p2, k)
Make a prediction, finding k nearest neighbours and summing them up
Parameters:
knn : knn matrix
p1 : feature 1 value
p2 : feature 2 value
k : sum k nearest neighbors
Returns: sum of k nearest neighbors
doKNN(kNN, k, skew, depth, price, long, short, store, feature1, feature2)
execute kNN filter
Parameters:
kNN : filter type
k : number k
skew : kNN minimum difference
depth : kNN matrix size limit
price : series
long : long condition
short : short condition
store : store the supplied features (if false, only checks the results without storage)
feature1 : feature 1 value
feature2 : feature 2 value
Returns: filter output
Machine Learning: kNN (New Approach)Description:
kNN is a very robust and simple method for data classification and prediction. It is very effective if the training data is large. However, it is distinguished by difficulty at determining its main parameter, K (a number of nearest neighbors), beforehand. The computation cost is also quite high because we need to compute distance of each instance to all training samples. Nevertheless, in algorithmic trading KNN is reported to perform on a par with such techniques as SVM and Random Forest. It is also widely used in the area of data science.
The input data is just a long series of prices over time without any particular features. The value to be predicted is just the next bar's price. The way that this problem is solved for both nearest neighbor techniques and for some other types of prediction algorithms is to create training records by taking, for instance, 10 consecutive prices and using the first 9 as predictor values and the 10th as the prediction value. Doing this way, given 100 data points in your time series you could create 10 different training records. It's possible to create even more training records than 10 by creating a new record starting at every data point. For instance, you could take the first 10 data points and create a record. Then you could take the 10 consecutive data points starting at the second data point, the 10 consecutive data points starting at the third data point, etc.
By default, shown are only 10 initial data points as predictor values and the 6th as the prediction value.
Here is a step-by-step workthrough on how to compute K nearest neighbors (KNN) algorithm for quantitative data:
1. Determine parameter K = number of nearest neighbors.
2. Calculate the distance between the instance and all the training samples. As we are dealing with one-dimensional distance, we simply take absolute value from the instance to value of x (| x – v |).
3. Rank the distance and determine nearest neighbors based on the K'th minimum distance.
4. Gather the values of the nearest neighbors.
5. Use average of nearest neighbors as the prediction value of the instance.
The original logic of the algorithm was slightly modified, and as a result at approx. N=17 the resulting curve nicely approximates that of the sma(20). See the description below. Beside the sma-like MA this algorithm also gives you a hint on the direction of the next bar move.
Esqvair's Neural Reversal Probability IndicatorIntroduction
Esqvair's Neural Reversal Probability Indicator is the indicator that shows probability of reversal.
Warning: This script should only be used on 1 minute chart.
How to use
When a signal appears (by default it is a green bar), a reversal should be expected.
The signal appears when the indicator value >= Threshold.
If you want more signals, you must lower the threshold, if less, you must increase the threshold.
For some assets, like Forex pairs, you have to optimize the threshold yourself, but for most stocks, the default threshold works well.
How well a threshold fits an asset depends on the volatility of the asset.
For most assets, the indicator ranges from 35 to 75.
Settings
Smoothing - The default is 1, which means no smoothing. Indicator smoothing by SMA.
Threshold - default 71.0 is responsible for the occurrence of signals, read "How to use" part to learn more
The Indicator
This indicator is a pre-trained neural network that was trained outside of TradingView and then its structure and weights values were converted to PineScript.
Warning: A neural network is a black box in the sense that although it can approximate any function, studying its structure will not give you any idea about the structure of the function being approximated.
Possible questions
Why does the indicator value most time range from 35 to 75 when the probability should ranges from 0 to 100?
-Due to some randomness in the markets, a neural network can never be 100% sure.
What data was used to train the neural network?
-This was BTCUSD 1 minute chart data from 02/05/2020 to 02/05/2022.
Where did you train the neural network and convert it to PineScript?
-I used a programming language that I know.
Tesla Coil MLThis is a re-implementation of @veryfid's wonderful Tesla Coil indicator to leverage basic Machine Learning Algorithms to help classify coil crossovers. The original Tesla Coil indicator requires extensive training and practice for the user to develop adequate intuition to interpret coil crossovers. The goal for this version is to help the user understand the underlying logic of the Tesla Coil indicator and provide a more intuitive way to interpret the indicator. The signals should be interpreted as suggestions rather than as a hard-coded set of rules.
NOTE: Please do NOT trade off the signals blindly. Always try to use your own intuition for understanding the coils and check for confluence with other indicators before initiating a trade.
Mean Shift Pivot ClusteringCore Concepts
According to Jeff Greenblatt in his book "Breakthrough Strategies for Predicting Any Market", Fibonacci and Lucas sequences are observed repeated in the bar counts from local pivot highs/lows. They occur from high to high, low to high, high to low, or low to high. Essentially, this phenomenon is observed repeatedly from any pivot points on any time frame. Greenblatt combines this observation with Elliott Waves to predict the price and time reversals. However, I am no Elliottician so it was not easy for me to use this in a practical manner. I decided to only use the bar count projections and ignore the price. I projected a subset of Fibonacci and Lucas sequences along with the Fibonacci ratios from each pivot point. As expected, a projection from each pivot point resulted in a large set of plotted data and looks like a huge gong show of lines. Surprisingly, I did notice clusters and have observed those clusters to be fairly accurate.
Fibonacci Sequence: 1, 2, 3, 5, 8, 13, 21, 34...
Lucas Sequence: 2, 1, 3, 4, 7, 11, 18, 29, 47...
Fibonacci Ratios (converted to whole numbers): 23, 38, 50, 61, 78, 127, 161...
Light Bulb Moment
My eyes may suck at grouping the lines together but what about clustering algorithms? I chose to use a gimped version of Mean Shift because it doesn't require me to know in advance how many lines to expect like K-Means. Mean shift is computationally expensive and with Pinescript's 500ms timeout, I had to make due without the KDE. In other words, I skipped the weighting part but I may try to incorporate it in the future. The code is from Harrison Kinsley . He's a fantastic teacher!
Usage
Search Radius: how far apart should the bars be before they are excluded from the cluster? Try to stick with a figure between 1-5. Too large a figure will give meaningless results.
Pivot Offset: looks left and right X number of bars for a pivot. Same setting as the default TradingView pivot high/low script.
Show Lines Back: show historical predicted lines. (These can change)
Use this script in conjunction with Fibonacci price retracement/extension levels and/or other support/resistance levels. If it's no where near a support/resistance and there's a projected time pivot coming up, it's probably a fake out.
Notes
Re-painting is intended. When a new pivot is found, it will project out the Fib/Lucas sequences so the algorithm will run again with additional information.
The script is for informational and educational purposes only.
Do not use this indicator by itself to trade!
WIPNNetworkLibrary "WIPNNetwork"
this is a work in progress (WIP) and prone to have some errors, so use at your own risk...
let me know if you find any issues..
Method for a generalized Neural Network.
network(x) Generalized Neural Network Method.
Parameters:
x : TODO: add parameter x description here
Returns: TODO: add what function returns
FunctionNNLayerLibrary "FunctionNNLayer"
Generalized Neural Network Layer method.
function(inputs, weights, n_nodes, activation_function, bias, alpha, scale) Generalized Layer.
Parameters:
inputs : float array, input values.
weights : float array, weight values.
n_nodes : int, number of nodes in layer.
activation_function : string, default='sigmoid', name of the activation function used.
bias : float, default=1.0, bias to pass into activation function.
alpha : float, default=na, if required to pass into activation function.
scale : float, default=na, if required to pass into activation function.
Returns: float
FunctionNNPerceptronLibrary "FunctionNNPerceptron"
Perceptron Function for Neural networks.
function(inputs, weights, bias, activation_function, alpha, scale) generalized perceptron node for Neural Networks.
Parameters:
inputs : float array, the inputs of the perceptron.
weights : float array, the weights for inputs.
bias : float, default=1.0, the default bias of the perceptron.
activation_function : string, default='sigmoid', activation function applied to the output.
alpha : float, default=na, if required for activation.
scale : float, default=na, if required for activation.
@outputs float
MLActivationFunctionsLibrary "MLActivationFunctions"
Activation functions for Neural networks.
binary_step(value) Basic threshold output classifier to activate/deactivate neuron.
Parameters:
value : float, value to process.
Returns: float
linear(value) Input is the same as output.
Parameters:
value : float, value to process.
Returns: float
sigmoid(value) Sigmoid or logistic function.
Parameters:
value : float, value to process.
Returns: float
sigmoid_derivative(value) Derivative of sigmoid function.
Parameters:
value : float, value to process.
Returns: float
tanh(value) Hyperbolic tangent function.
Parameters:
value : float, value to process.
Returns: float
tanh_derivative(value) Hyperbolic tangent function derivative.
Parameters:
value : float, value to process.
Returns: float
relu(value) Rectified linear unit (RELU) function.
Parameters:
value : float, value to process.
Returns: float
relu_derivative(value) RELU function derivative.
Parameters:
value : float, value to process.
Returns: float
leaky_relu(value) Leaky RELU function.
Parameters:
value : float, value to process.
Returns: float
leaky_relu_derivative(value) Leaky RELU function derivative.
Parameters:
value : float, value to process.
Returns: float
relu6(value) RELU-6 function.
Parameters:
value : float, value to process.
Returns: float
softmax(value) Softmax function.
Parameters:
value : float array, values to process.
Returns: float
softplus(value) Softplus function.
Parameters:
value : float, value to process.
Returns: float
softsign(value) Softsign function.
Parameters:
value : float, value to process.
Returns: float
elu(value, alpha) Exponential Linear Unit (ELU) function.
Parameters:
value : float, value to process.
alpha : float, default=1.0, predefined constant, controls the value to which an ELU saturates for negative net inputs. .
Returns: float
selu(value, alpha, scale) Scaled Exponential Linear Unit (SELU) function.
Parameters:
value : float, value to process.
alpha : float, default=1.67326324, predefined constant, controls the value to which an SELU saturates for negative net inputs. .
scale : float, default=1.05070098, predefined constant.
Returns: float
exponential(value) Pointer to math.exp() function.
Parameters:
value : float, value to process.
Returns: float
function(name, value, alpha, scale) Activation function.
Parameters:
name : string, name of activation function.
value : float, value to process.
alpha : float, default=na, if required.
scale : float, default=na, if required.
Returns: float
derivative(name, value, alpha, scale) Derivative Activation function.
Parameters:
name : string, name of activation function.
value : float, value to process.
alpha : float, default=na, if required.
scale : float, default=na, if required.
Returns: float
MLLossFunctionsLibrary "MLLossFunctions"
Methods for Loss functions.
mse(expects, predicts) Mean Squared Error (MSE) " MSE = 1/N * sum ((y - y')^2) ".
Parameters:
expects : float array, expected values.
predicts : float array, prediction values.
Returns: float
binary_cross_entropy(expects, predicts) Binary Cross-Entropy Loss (log).
Parameters:
expects : float array, expected values.
predicts : float array, prediction values.
Returns: float
Financial Astrology Indexes ML Daily TrendDaily trend indicator based on financial astrology cycles detected with advanced machine learning techniques for some of the most important market indexes: DJI, UK100, SPX, IBC, IXIC, NI225, BANKNIFTY, NIFTY and GLD fund (not index) for Gold predictions. The daily price trend is forecasted through planets cycles (angular aspects, speed phases, declination zone), fast cycles are based on Moon, Mercury, Venus and Sun and Mid term cycles are based on Mars, Vesta and Ceres . The combination of all this cycles produce a daily price trend prediction that is encoded into a PineScript array using binary format "0 or 1" that represent sell and buy signals respectively. The indicator provides signals since 2021-01-01 to 2022-12-31, the past months signals purpose is to support backtesting of the indicator combined with other technical indicator entries like MAs, RSI or Stochastic . For future predictions besides 2022 a machine learning models re-train phase will be required.
When the signal moving average is increasing from 0 to 1 indicates an increase of buy force, when is decreasing from 1 to 0 indicates an increase in sell force, finally, when is sideways around the 0.4-0.6 area predicts a period of buy/sell forces equilibrium, traders indecision which result in a price congestion within a narrow price range.
We also have published same indicator for Crypto-Currencies research portfolio:
DISCLAIMER: This indicator is experimental and don’t provide financial or investment advice, the main purpose is to demonstrate the predictive power of financial astrology. Any allocation of funds following the documented machine learning model prediction is a high-risk endeavour and it’s the users responsibility to practice healthy risk management according to your situation.
Financial Astrology Crypto ML Daily TrendThis daily trend indicator is based on financial astrology cycles detected with advanced machine learning techniques for the crypto-currencies research portfolio: ADA, BAT, BNB, BTC, DASH, EOS, ETC, ETH, LINK, LTC, XLM, XMR, XRP, ZEC and ZRX. The daily price trend is forecasted through this planets cycles (angular aspects, speed, declination), fast ones are based on Moon, Mercury, Venus and Sun and Mid term cycles are based on Mars, Vesta and Ceres. The combination of all this cycles produce a daily price trend prediction that is encoded into a PineScript array using binary format "0 or 1" that represent sell and buy signals respectively. The indicator provides signals since 2021-01-01 to 2022-12-31, the past months signals purpose is to support backtesting of the indicator combined with other technical indicator entries like MAs, RSI or Stochastic. For future predictions besides 2022 a machine learning models re-train phase will be required.
The resolution of this indicator is 1D, you can tune a parameter where you can determine how many future bars of daily trend are plotted and adjust an hours shift to anticipate future signals into current bar in order to produce a leading indicator effect to anticipate the trend changes with some hours of anticipation. Combined with technical analysis indicators this daily trend is very powerful because can help to produce approximately 60% of profitable signals based on the backtesting results. You can look at our open source Github repositories to validate accuracy using the backtesting strategies we have implemented in Jesse Crypto Trading Framework as proof of concept of the predictive potential of this indicator. Alternatively, we have implemented a PineScript strategy that use this indicator, just consider that we are pending to do signals update to the period July 2021 to December 2022: This strategy have accumulated more than 110 likes and many traders have validated the predictive power of Financial Astrology.
DISCLAIMER: This indicator is experimental and don’t provide financial or investment advice, the main purpose is to demonstrate the predictive power of financial astrology. Any allocation of funds following the documented machine learning model prediction is a high-risk endeavour and it’s the users responsibility to practice healthy risk management according to your situation.
VWMA with kNN Machine Learning: MFI/ADXThis is an experimental strategy that uses a Volume-weighted MA (VWMA) crossing together with Machine Learning kNN filter that uses ADX and MFI to predict, whether the signal is useful. k-nearest neighbours (kNN) is one of the simplest Machine Learning classification algorithms: it puts input parameters in a multidimensional space, and then when a new set of parameters are given, it makes a prediction based on plurality vote of its k neighbours.
Money Flow Index (MFI) is an oscillator similar to RSI, but with volume taken into account. Average Directional Index (ADX) is an indicator of trend strength. By putting them together on two-dimensional space and checking, whether nearby values have indicated a strong uptrend or downtrend, we hope to filter out bad signals from the MA crossing strategy.
This is an experiment, so any feedback would be appreciated. It was tested on BTC/USDT pair on 5 minute timeframe. I am planning to expand this strategy in the future to include more moving averages and filters.
Morun Astro Trend MAs cross StrategyAstrology machine learning cycles indicator signals with technical MAs indicators strategy, based on signals index of Github project github.com
Machine Learning: kNN-based Strategy (update)kNN-based Strategy (FX and Crypto)
Description:
This update to the popular kNN-based strategy features:
improvements in the business logic,
an adjustible k value for the kNN model,
one more feature (MOM),
a streamlined signal filter and
some other minor fixes.
Now this script works in all timeframes !
I intentionally decided to publish this script separately
in order for the users to see the differences.
Machine Learning: LVQ-based StrategyLVQ-based Strategy (FX and Crypto)
Description:
Learning Vector Quantization (LVQ) can be understood as a special case of an artificial neural network, more precisely, it applies a winner-take-all learning-based approach. It is based on prototype supervised learning classification task and trains its weights through a competitive learning algorithm.
Algorithm:
Initialize weights
Train for 1 to N number of epochs
- Select a training example
- Compute the winning vector
- Update the winning vector
Classify test sample
The LVQ algorithm offers a framework to test various indicators easily to see if they have got any *predictive value*. One can easily add cog, wpr and others.
Note: TradingViews's playback feature helps to see this strategy in action. The algo is tested with BTCUSD/1Hour.
Warning: This is a preliminary version! Signals ARE repainting.
***Warning***: Signals LARGELY depend on hyperparams (lrate and epochs).
Style tags: Trend Following, Trend Analysis
Asset class: Equities, Futures, ETFs, Currencies and Commodities
Dataset: FX Minutes/Hours+++/Days
Machine Learning: Logistic RegressionMulti-timeframe Strategy based on Logistic Regression algorithm
Description:
This strategy uses a classic machine learning algorithm that came from statistics - Logistic Regression (LR).
The first and most important thing about logistic regression is that it is not a 'Regression' but a 'Classification' algorithm. The name itself is somewhat misleading. Regression gives a continuous numeric output but most of the time we need the output in classes (i.e. categorical, discrete). For example, we want to classify emails into “spam” or 'not spam', classify treatment into “success” or 'failure', classify statement into “right” or 'wrong', classify election data into 'fraudulent vote' or 'non-fraudulent vote', classify market move into 'long' or 'short' and so on. These are the examples of logistic regression having a binary output (also called dichotomous).
You can also think of logistic regression as a special case of linear regression when the outcome variable is categorical, where we are using log of odds as dependent variable. In simple words, it predicts the probability of occurrence of an event by fitting data to a logit function.
Basically, the theory behind Logistic Regression is very similar to the one from Linear Regression, where we seek to draw a best-fitting line over data points, but in Logistic Regression, we don’t directly fit a straight line to our data like in linear regression. Instead, we fit a S shaped curve, called Sigmoid, to our observations, that best SEPARATES data points. Technically speaking, the main goal of building the model is to find the parameters (weights) using gradient descent.
In this script the LR algorithm is retrained on each new bar trying to classify it into one of the two categories. This is done via the logistic_regression function by updating the weights w in the loop that continues for iterations number of times. In the end the weights are passed through the sigmoid function, yielding a prediction.
Mind that some assets require to modify the script's input parameters. For instance, when used with BTCUSD and USDJPY, the 'Normalization Lookback' parameter should be set down to 4 (2,...,5..), and optionally the 'Use Price Data for Signal Generation?' parameter should be checked. The defaults were tested with EURUSD.
Note: TradingViews's playback feature helps to see this strategy in action.
Warning: Signals ARE repainting.
Style tags: Trend Following, Trend Analysis
Asset class: Equities, Futures, ETFs, Currencies and Commodities
Dataset: FX Minutes/Hours/Days
Machine Learning: Perceptron-based strategyPerceptron-based strategy
Description:
The Learning Perceptron is the simplest possible artificial neural network (ANN), consisting of just a single neuron and capable of learning a certain class of binary classification problems. The idea behind ANNs is that by selecting good values for the weight parameters (and the bias), the ANN can model the relationships between the inputs and some target.
Generally, ANN neurons receive a number of inputs, weight each of those inputs, sum the weights, and then transform that sum using a special function called an activation function. The output of that activation function is then either used as the prediction (in a single neuron model) or is combined with the outputs of other neurons for further use in more complex models.
The purpose of the activation function is to take the input signal (that’s the weighted sum of the inputs and the bias) and turn it into an output signal. Think of this activation function as firing (activating) the neuron when it returns 1, and doing nothing when it returns 0. This sort of computation is accomplished with a function called step function: f(z) = {1 if z > 0 else 0}. This function then transforms any weighted sum of the inputs and converts it into a binary output (either 1 or 0). The trick to making this useful is finding (learning) a set of weights that lead to good predictions using this activation function.
Training our perceptron is simply a matter of initializing the weights to zero (or random value) and then implementing the perceptron learning rule, which just updates the weights based on the error of each observation with the current weights. This has the effect of moving the classifier’s decision boundary in the direction that would have helped it classify the last observation correctly. This is achieved via a for loop which iterates over each observation, making a prediction of each observation, calculating the error of that prediction and then updating the weights accordingly. In this way, weights are gradually updated until they converge. Each sweep through the training data is called an epoch.
In this script the perceptron is retrained on each new bar trying to classify this bar by drawing the moving average curve above or below the bar.
This script was tested with BTCUSD, USDJPY, and EURUSD.
Note: TradingViews's playback feature helps to see this strategy in action.
Warning: Signals ARE repainting.
Style tags: Trend Following, Trend Analysis
Asset class: Equities, Futures, ETFs, Currencies and Commodities
Dataset: FX Minutes/Hours+/Days
Machine Learning: kNN-based Strategy (mtf)This is a multi-timeframe version of the kNN-based strategy.
Machine Learning: kNN-based StrategykNN-based Strategy (FX and Crypto)
Description:
This strategy uses a classic machine learning algorithm - k Nearest Neighbours (kNN) - to let you find a prediction for the next (tomorrow's, next month's, etc.) market move. Being an unsupervised machine learning algorithm, kNN is one of the most simple learning algorithms.
To do a prediction of the next market move, the kNN algorithm uses the historic data, collected in 3 arrays - feature1, feature2 and directions, - and finds the k-nearest
neighbours of the current indicator(s) values.
The two dimensional kNN algorithm just has a look on what has happened in the past when the two indicators had a similar level. It then looks at the k nearest neighbours,
sees their state and thus classifies the current point.
The kNN algorithm offers a framework to test all kinds of indicators easily to see if they have got any *predictive value*. One can easily add cog, wpr and others.
Note: TradingViews's playback feature helps to see this strategy in action.
Warning: Signals ARE repainting.
Style tags: Trend Following, Trend Analysis
Asset class: Equities, Futures, ETFs, Currencies and Commodities
Dataset: FX Minutes/Hours+++/Days
