Multifactor research: check scale and overlap before combining scores
Combine fictional normalized scores while preserving direction, timing and validation rules.
Specify what is being combined
Multifactor research can combine company characteristics or explain returns with factor-return regressions. This article concerns normalized security scores and a weighted composite.
More features do not guarantee more information. Identify each feature’s meaning and availability before combining it.
A fictional weighted score
EX-A has scores 0.8 and 0.2; EX-B has 0.6 and 0.7. Higher values are preferred by definition. Equal weights produce 0.50 for A and 0.65 for B.
Weights of 0.8 and 0.2 produce 0.68 for A and 0.62 for B, reversing the ranking. These fictional weights demonstrate sensitivity, not a recommended setting.
Align scale and direction
Adding yen amounts to percentages directly can let units dominate. Specify normalization, ties, eligible observations and direction. Ranking can remove information about the magnitude of differences.
Predefine whether a missing feature excludes a security or triggers weight redistribution. Redistribution can change what the score means across securities.
Inspect overlap and selection
Highly correlated scores may repeat information. Examine correlations over time and sector or size concentrations. Low correlation alone does not prove predictive power.
A period used to select features or weights is not an untouched evaluation period. Record the selection process and the number of candidate definitions considered.
Understand the public composite example
KabuForge’s public course combines momentum features from different windows under fixed weights. This does not establish automatic value or quality inputs or optimal weights.
Diagnose individual features first. Read rank normalization, factor correlation and out-of-sample evaluation.
Public sources and limitations
Concept references: Fama–French five-factor definitions. Additional reference: definitions of multiple factor-return series. Software workflow and limitations: KabuForge public GUI research course. Reviewed October 6, 2026. The fictional examples are teaching calculations, not course execution results.
The public course is research-only and does not certify historical information availability. A concept explanation does not establish that every method is implemented in the software. This article recommends no securities, trades or settings. Read the disclaimer.
Continue this topic
Applies to: main source candidate (unpublished) · Public source docs/en_US/GUI_RESEARCH_COURSES.md
Reviewed · KabuForge