2005issue C041-4
Hybrid decision trees and pattern recognition for trend rules
Unaided decision quality can rise as information is first added and then fall once information saturation is reached. A hybrid-decision-system extracts patterned instances through expert-style rules, trains a pattern-recognition network to a designated output, and can keep a decision-tree backbone while a trend-following procedure stays with a primary-trend until the opposite direction is confirmed.
- Unaided decision quality can rise as information is first added and then fall once information saturation is reached.
- An expert-style decision tree stores knowledge-forms as facts, procedures, or rules and can form a certainty-factor from successive levels, but the rule set does not learn or adapt.
- Pattern recognition begins without stored knowledge, updates when new examples are shown, and in the hybrid-decision-system trains on instances and attributes extracted by those rules.
- Trend following identifies a primary-trend that stays in force until the opposite direction is confirmed and treats corrections as secondary to that direction.
When added information starts to hurt
Unaided decision quality can rise as information is first added and then fall once a saturation point is reached. That turning point is information saturation: adding further observations stops improving unaided decision quality and begins to degrade it.
A rule tree that stores knowledge but does not learn
An expert-style system stores knowledge as facts, procedures, or rules and activates those rules at successive levels of a decision tree. Those encodings are the knowledge-forms the rule base can accept.
When period-specific facts are supplied, an expert-style system can combine level-wise probabilities into a certainty-factor. The rule set itself does not learn or adapt.
Pattern recognition without a stored rule base
A pattern-recognition network begins without stored knowledge, updates when shown new examples, and can shift emphasis from older associations to newer ones. Pattern recognition is positioned for classification when specified influences map to known outcomes that need not be linear or already modeled.
Building the hybrid decision system
In the hybrid construction, expert-system rules extract patterned instances and attributes, the data are stored and preprocessed, and a network then trains to a designated output. That construction is the hybrid-decision-system.
Tree links from the rule base can initialize the network so the combined model keeps an explicit logical backbone while remaining able to operate on incomplete observations.
Trend following as a holding-period procedure
A trend-following procedure identifies a primary bull or bear direction that remains in force until the opposite direction is confirmed, treating range-bound stretches as narrower pauses rather than as an absence of trend. That underlying bull or bear direction is the primary-trend.
Corrective moves may last weeks or months without ending the primary-trend. The stated task is to detect that trend change rather than to forecast a future price level.
After a primary-trend is identified, the described procedure seeks early participation in the large directional move and treats counter-trend pullbacks as points at which exposure may be increased rather than abandoned.
An editorial reading of the combined process
Editorial: Once the decision tree has frozen entry, exit, and abstention branches, pattern recognition is used to score which ordered observations fire those branches. The trend-following holding period stays with the primary-trend until the opposite direction is confirmed. The combined procedure is evaluated only against a stated out-of-sample baseline.
All readings on this track · 25 readings
- 1986Construct a decision procedure that revises itself
- 1989Finish the volume checklist before scoring the breakout
- 1989Constructing supervised forecasts on moving averages
- 1991Candlestick labels as stacked construction tests
- 1992Walk-forward evaluation of weekly price-change patterns
- 1993RSI price pattern templates and open interest
- 1994Constructing a dual-net day-ahead index direction forecast
- 1994A clocked stochastic second crest with a window-high stop
- 1996Volatility-ratio, inside-day and narrow-range-4 entry construction
- 1998Sliding-window correlation for cup-and-handle construction
- 2000Constructing rectangles for breakout hypotheses
- 2001Turning one candle into a ranked numeric object
- 2002Fuzzy-scored chart patterns as testable rules
- 2002From hot-zones to an open-close-matrix
- 2003Volume pressure and a band-clearing breakout case
- 2004Evaluating chart patterns against price objectives
- 2004Cobweb turning points from price structure
- 2005Hybrid decision trees and pattern recognition for trend rules
- 2005Two-bar zone codes for testable pattern systems
- 2005Price bar pattern construction and next-bar frequency
- 2008Observe markets before following pattern or system rules
- 2012Treat a four-leg Fibonacci completion as an unpaid hypothesis
- 2014Hidden three-channel regression signals for stock and call option entries
- 2014A shared daily-chart-level framework for session trades and swing holds
- 2015Condensed candlestick signatures