Method of rule constrained statistical price formation recognition
A method of developing a rule-constrained statistical pattern recognizer applicable to price formation recognition includes assembling input data containing examples of patterns to be recognized and establishing mandatory recognition rules. The recognition rules are programmed to construct an underspecified or underconstrained recognition model which is applied to the assembled data to produce candidate patterns. The candidate patterns are reviewed and identified as valid or invalid and for each pattern type a residual statistical model is produced based on the candidate patterns identified as valid. A filter is used to ensure that custom conditions such as duration relationships, height relationships and volume requirements are met.
1. A rule constrained statistical recognition system for determining rated pattern candidates from time-series data, the system comprising:
a processor;
a rule constrained recognition module for identifying pattern candidates, corresponding to a predetermined pattern, in the time-series data, using under-constrained recognition models to include all true positives while minimizing or eliminating false negatives;
a residual statistical module for identifying each pattern candidate as valid or invalid using residual statistical models and for rating each valid pattern candidate;
storage means for storing the rated pattern candidates; and
a user interface for displaying the rated pattern candidates.
2. The system of claim 1 , further comprising a characterization module for identifying characteristics of each pattern candidate and wherein the storage means comprises means for storing pattern candidates and each candidate's identified characteristics.
3. The system of claim 2 , further comprising means for conducting technical analysis on the time-series data.
4. The system of claim 2 , further comprising a filter for excluding pattern candidates based on the identified characteristics.
5. The system of claim 2 , further comprising a filter for validating pattern candidates based on the identified characteristics.
6. The system of claim 1 , wherein the storage means stores each rated pattern candidate that is outside a feasibility threshold limit for retraining of the residual statistical module.
7. The system of claim 1 , wherein the residual statistical module is a radial basis function neural net module.
8. The system of claim 1 , further comprising at least one of:
a training module, a genetic algorithm parameter tuning module, a genetic algorithm feature selection module, and a Bayesian regularizer module for training the rule constrained recognition module or the residual statistical module.
9. A computer-implemented method of rule constrained statistical pattern recognition for determination of predetermined patterns in time-series data, the method comprising:
using an under-constrained rule constrained recognition model to identify pattern candidates by including all true positives while minimizing or eliminating false negatives;
identifying each pattern candidate as valid or invalid using residual statistical models;
assigning a rating to each valid pattern candidate;
determining the characteristics of each pattern candidate;
filtering the resulting rated pattern candidates to exclude pattern candidates based on the pattern candidate's characteristics; and
displaying the rated pattern candidates on a user interface.
10. A computer-implemented method of training a rule-constrained statistical pattern recognizer, the method comprising:
assembling data for input space of recognition model, the data including examples of pattern types to be recognized;
establishing recognition rules;
generating, using the recognition rules, an under-specified rule-constrained recognition model to include all true positives while minimizing or eliminating false negatives;
applying the rule-constrained recognition model to the assembled data to produce candidate patterns;
identifying features of candidate patterns for use in a statistical model to assess validity of different candidate types;
generating, for each pattern type, a residual statistical model based on the candidate pattern identified as valid; and
training the rule-constrained statistical pattern recognizer using the generated rule constrained recognition model and the residual statistical model to recognize price formation.
11. The method of claim 10 , wherein the recognition rules are mandatory rules.
12. The method of claim 10 , wherein the step of identifying candidate patterns comprises labeling each candidate pattern as being valid or invalid.
13. The method of claim 10 , further comprising applying a filter to each pattern type.
14. The method of claim 13 , wherein applying the filter comprises ensuring that duration relationships are respected.
15. The method of claim 13 , wherein applying the filter comprises ensuring that height relationships are respected.
16. The method of claim 13 , wherein applying the filter comprises ensuring that volume requirements are met.
17. A rule constrained statistical recognition system for determining rated pattern candidates from data, the system comprising:
a processor;
a rule constrained recognition module for identifying pattern candidates, corresponding to a predetermined pattern, in the data, using under-constrained recognition models to include all true positives while minimizing or eliminating false negatives;
a residual statistical module for identifying each pattern candidate as valid or invalid using residual statistical models and for rating each valid pattern candidate;
storage means for storing the rated pattern candidates; and
a user interface for displaying the rated pattern candidates.