Sniffer
╭━━━╮╱╱╱╱╭━╮╭━╮
┃╭━╮┃╱╱╱╱┃╭╯┃╭╯
┃╰━━┳━╮╭┳╯╰┳╯╰┳━━┳━╮
╰━━╮┃╭╮╋╋╮╭┻╮╭┫┃━┫╭╯
┃╰━╯┃┃┃┃┃┃┃╱┃┃┃┃━┫┃
╰━━━┻╯╰┻╯╰╯╱╰╯╰━━┻╯
Overview
A vast majority of modern data analysis & modelling techniques rely upon the idea of hidden patterns, wether it is some type of visualisation tool or some form of a complex machine learning algorithm, the one thing that they have in common is the belief, that patterns tell us what’s hidden behind plain numbers. The same philosophy has been adopted by many traders & investors worldwide, there’s an entire school of thought that operates purely based on chart patterns. This is where Sniffer comes in, it is a tool designed to simplify & quantify the job of pattern recognition on any given price chart, by combining various factors & techniques that generate high-quality results.
This tool analyses bars selected by the user, and highlights bar clusters on the chart that exhibit similar behaviour across multiple dimensions. It can detect a single candle pattern like hammers or dojis, or it can handle multiple candles like morning/evening stars or double tops/bottoms, and many more. In fact, the tool is completely independent of such specific candle formations, instead, it works on the idea of vector similarity and generates a degree of similarity for every single combination of candles. Only the top-n matches are highlighted, users get to choose which patterns they want to analyse and to what degree, by customising the feature-space.
Background
In the world of trading, a common use-case is to scan a price chart for some specific candlestick formations & price structures, and then the chart is further analysed in reference to these events. Traders are often trying to answer questions like, when was the last time price showed similar behaviour, what are the instances similar to what price is doing right now, what happens when price forms a pattern like this, what were some of other indicators doing when this happened last(RSI, CCI, ADX etc), and many other abstract ideas to have a stronger confluence or to confirm a bias.Having such a context can be vital in making better informed decisions, but doing this manually on a chart that has thousands of candles can have many disadvantages. It’s tedious, human errors are rather likely, and even if it’s done with pin-point accuracy, chances are that we’ll miss out on many pieces of information. This is the thought that gave birth to Sniffer .
Sniffer tries to provide a general solution for pattern-based analysis by deploying vector-similarity computation techniques, that cover the full-breadth of a price chart and generate a list of top-n matches based on the criteria selected by the user. Most of these techniques come from the data science space, where vector similarity is often implemented to solve classification & clustering problems. Sniffer uses same principles of vector comparison, and computes a degree of similarity for every single candle formation within the selected range, and as a result generates a similarity matrix that captures how similar or dissimilar a set of candles is to the input set selected by the user.
How It Works
A brief overview of how the tool is implemented:
- Every bar is processed, and a set of features are mapped to it.
- Bars selected by the user are captured, and saved for later use.
- Once the all the bars have been processed, candles are back-tracked and degree of similarity is computed for every single bar(max-limit is 5000 bars).
- Degree of similarity is computed by comparing attributes like price range, candle breadth & volume etc.
- Similarity matrix is sorted and top-n results are highlighted on the chart through boxes of different colors.
A brief overview of the features space for bars:
- Range: Difference between high & low
- Body: Difference between close & open
- Volume: Traded volume for that candle
- Head: Upper wick for green candles & lower wick for red candles
- Tail: Lower wick for green candles & upper wick for red candles
- BTR: Body to Range ratio
- HTR: Head to Range ratio
- TTR: Tail to Range ratio
- HTB: Head to Body ratio
- TTB: Tail to Body ratio
- ROC: Rate of change for HL2 for four different periods
- RSI: Relative Strength Index
- CCI: Commodity Channel Index
- Stochastic: Stochastic Index
- ADX: DMI+, DMI- & ADX
A brief overview of how degree of similarity is calculated:
- Each bar set is compared to the inout bar set within the selected feature space
- Features are represented as vectors, and distance between the vectors is calculated
- Shorter the distance, greater the similarity
- Different distance calculation methods are available to choose from, such as Cosine, Euclidean, Lorentzian, Manhattan, & Pearson
- Each method is likely to generate slightly different results, users are expected to select the method & the feature space that best fits their use-case
How To Use It
- Usage of this tool is relatively straightforward, users can add this indicator to their chart and similar clusters will be highlighted automatically
- Users need to select a time range that will be treated as input, and bars within that range become the input formation for similarity calculations
- Boxes will be draw around the clusters that fit the matching criteria
- Boxes are color-coded, green color boxes represent the top one-third of the top-n matches, yellow boxes represent the middle third, red boxes are for bottom third, and white box represents user-input
- Boxes colors will be adjusted as you adjust input parameters, such as number of matches or look-back period
User Settings
Users can configure the following options:
- Select the time-range to set input bars
- Select the look-back period, number of candles to backtrack for similarity search
- Select the number of top-n matches to show on the chart
- Select the method for similarity calculation
- Adjust the feature space, this enables addition of custom features, such as pattern recognition, technical indicators, rate of change etc
- Toggle verbosity, shows degree of similarity as a percentage value inside the box
Top Features
- Pattern Agnostic: Designed to work with variable number of candles & complex patterns
- Customisable Feature Space: Users get to add custom features to each bar
- Comprehensive Comparison: Generates a degree of similarity for all possible combinations
Final Note
- Similarity matches will be shown only within last 4500 bars.
- In theory, it is possible to compute similarity for any size candle formations, indicator has been tested with formations of 50+ candles, but it is recommended to select smaller range for faster & cleaner results.
- As you move to smaller time frames, selected time range will provide a larger number of candles as input, which can produce undesired results, it is advised to adjust your selection when you change time frames. Seeking suggestions on how to directly receive bars as user input, instead of time range.
- At times, users may see array index out of bound error when setting up this indicator, this generally happens when the input range is not properly configured. So, it should disappear after you select the input range, still trying to figure out where it is coming from, suggestions are welcome.
Credits
- @HeWhoMustNotBeNamed for publishing such a handy PineScript Logger, it certainly made the job a lot easier.
Patternrecognition
FunctionPatternDecompositionLibrary "FunctionPatternDecomposition"
Methods for decomposing price into common grid/matrix patterns.
series_to_array(source, length) Helper for converting series to array.
Parameters:
source : float, data series.
length : int, size.
Returns: float array.
smooth_data_2d(data, rate) Smooth data sample into 2d points.
Parameters:
data : float array, source data.
rate : float, default=0.25, the rate of smoothness to apply.
Returns: tuple with 2 float arrays.
thin_points(data_x, data_y, rate) Thin the number of points.
Parameters:
data_x : float array, points x value.
data_y : float array, points y value.
rate : float, default=2.0, minimum threshold rate of sample stdev to accept points.
Returns: tuple with 2 float arrays.
extract_point_direction(data_x, data_y) Extract the direction each point faces.
Parameters:
data_x : float array, points x value.
data_y : float array, points y value.
Returns: float array.
find_corners(data_x, data_y, rate) ...
Parameters:
data_x : float array, points x value.
data_y : float array, points y value.
rate : float, minimum threshold rate of data y stdev.
Returns: tuple with 2 float arrays.
grid_coordinates(data_x, data_y, m_size) transforms points data to a constrained sized matrix format.
Parameters:
data_x : float array, points x value.
data_y : float array, points y value.
m_size : int, default=10, size of the matrix.
Returns: flat 2d pseudo matrix.
Pattern Recognition Probabilities [racer8]Brief 🌟
Pattern Recognition Probabilities (PRP) is a REALLY smart indicator. It uses the correlation coefficient formula to determine if the current set of bars resembles that of past patterns. It counts the number of times the current pattern has occurred in the past and looks at how it performed historically to determine the probability of an up move, down move, or neutral move.
I'd like to say, I'm proud of this indicator 😆🤙 This is the SMARTEST indicator I have ever made 🧠🧠🧠
Note: PRP doesn't give you actual probabilities, but gives you instead the historical occurrences of up, down, and neutral moves that resulted after the pattern. So you can calculate probabilities based on these valuable statistics. So for example, PRP can tell you this pattern has historically resulted in 55 up moves, 20 down moves, and 60 neutral moves.
Parameters 🌟
You can adjust the Pattern length, Minimum correlation, Statistics lookback, Exit after time, and Atr multiplier parameters.
Pattern length - determines how long the pattern is
Minimum correlation - determines the minimum correlation coefficient needed to pass as a similiar enough pattern.
Statistics lookback - lookback period for gathering all the patterns in the past.
Exit after time - determines when exit occurred (number of periods after pattern) ; is the point that represents the pattern's result.
Atr multiplier - determines minimum atr move needed to qualify whether result was an up/down move or a neutral move. If a particular historical pattern resulted in a move that was less than the min atr, then it is recorded as a neutral move in the statistics.
Thanks for reading! 🙏
Good luck 🍀 Stay safe 😷 Drink lots of water💧
Enjoy! 🥳 and Hit the like button! 👍
Test: Pattern RecognitionEXPERIMENTAL:
a test on how to compare price at different frequency's with static patterns.
