Trading Agent Dev Log #8 — Multi-Factor Weighted Score System and Composite Score Migration

Thu, 02 Apr 2026 00:00:00 +0900

Overview

Previous Post: Trading Agent Dev Log #7 covered agent settings UI and signal card improvements. In this #8 installment, the single min_rr_score gate was replaced with a 5-factor Composite Score system, along with building a new stock research page, sell validation logic, and project rebranding — a major overhaul spanning 41 commits.

1. Stock Research Page (Stock Info)

Problem

Even when a signal was generated, there was no page to view the stock’s fundamental analysis, technical indicators, and institutional flow at a glance. Having to open an external brokerage HTS every time caused delays in decision-making.

Implementation

A /api/research router was added to the backend with 5 endpoints (basic info, financials, technical indicators, institutional flow, news/disclosures). The frontend was split into 9 section components.

// frontend/src/components/stockinfo/ structure
DiscoverySidebar.tsx // Stock search sidebar
ResearchHeader.tsx // Stock basic info header
PriceChartSection.tsx // Candlestick chart + technical indicators
FundamentalsSection.tsx // Key financial metrics
ValuationSection.tsx // Valuation comparison
InvestorFlowSection.tsx // Foreign/institutional flow
PeerSection.tsx // Industry peer comparison
InsiderSection.tsx // Insider trading
SignalHistorySection.tsx // Past signal history

Charts were upgraded from simple line charts to candlestick + volume + moving averages (MA) + Bollinger Bands (BB) overlays, and technical indicators are displayed as mini-chart 2x2 grid cards for RSI, MACD, Bollinger Bands, and volume trends respectively.

2. Signal Pipeline Improvements — Linear Confidence and Sell Validation

From Sigmoid to Linear Mapping

The existing sigmoid-based compute_confidence had a “dead zone” where R/R scores between 0.5 and 1.5 produced nearly identical confidence values. This was replaced with linear mapping and the min_rr_score threshold was lowered to 0.3 to capture a wider range of signals.

Sell (SELL) Validation Logic

A problem was discovered where SELL signals were generated for stocks not held in the portfolio. Two hard gates were added:

Risk Manager: Reject SELL for unheld stocks + minimum hold time validation
Market Scanner: Force-convert SELL to HOLD for unheld stocks

SELL/HOLD direction rules were also explicitly added to the expert panel prompts so that the Chief Analyst gives opinions while being aware of the current holdings.

3. Multi-Factor Composite Score System

This is the core change of this installment. Signal filtering that relied on a single R/R score was completely replaced with a weighted sum of 5 independent factors.

flowchart LR
 subgraph 5-Factors["5 Sub-Scores"]
 A["R/R Ratio<br/>(Risk-Reward)"]
 B["Expert Consensus<br/>(Agreement Level)"]
 C["Fundamental<br/>(PER, ROE, etc.)"]
 D["Technical Momentum<br/>(RSI, MACD, Volume)"]
 E["Institutional Flow<br/>(Foreign/Institutional)"]
 end

 subgraph Weights["Weight Normalization"]
 W["normalize_weights()"]
 end

 subgraph Quality["Data Quality"]
 Q["confidence_grades<br/>A=1.0 B=0.85 C=0.6 D=0.3"]
 end

 A --> W
 B --> W
 C --> W
 D --> W
 E --> W
 W -->|"weighted sum"| R["raw score"]
 R -->|"x quality"| F["Composite Score<br/>(0~100)"]
 Q --> F

Sub-Score Design

Each factor is normalized to a 0-1 range, returning a default of 0.5 (neutral) when data is unavailable.

# backend/app/models/composite_score.py

def score_fundamental(
 per: float | None = None,
 roe: float | None = None,
 debt_ratio: float | None = None,
 operating_margin: float | None = None,
) -> float:
 """Normalize each metric independently to 0-1 and return the average.
 Missing metrics are excluded from the calculation."""
 components: list[float] = []
 if per is not None and per > 0:
 components.append(min(max(1.0 - per / 40.0, 0.0), 1.0))
 if roe is not None:
 components.append(min(max(roe / 30.0, 0.0), 1.0))
 # ... debt_ratio, operating_margin follow the same pattern
 return sum(components) / len(components) if components else 0.5

The institutional/foreign flow score uses sigmoid normalization. The net purchase total is divided by a base amount (default 1 billion KRW) and mapped to the -1 to 1 range.

def score_institutional_flow(
 foreign_net: float = 0,
 institution_net: float = 0,
 scale: float = 1_000_000_000, # 1 billion KRW
) -> float:
 combined = foreign_net + institution_net
 return 1.0 / (1.0 + math.exp(-combined / scale))

Weight Normalization and Aggregation

User-configured weights are automatically normalized to sum to 1.0. The final score is computed by multiplying the weighted sum by a data quality multiplier and converting to a 0-100 scale.

def compute_composite_score(
 rr_score: float,
 calibration_ceiling: float = 2.0,
 expert_analyses: list[dict] | None = None,
 dart_financials: dict | None = None,
 technicals: dict | None = None,
 investor_trend: dict | None = None,
 confidence_grades: dict[str, str] | None = None,
 weights: dict[str, float] | None = None,
) -> float:
 w = normalize_weights(weights) if weights else dict(DEFAULT_WEIGHTS)
 # ... compute 5 sub-scores ...
 raw = (
 w["rr_ratio"] * rr_sub
 + w["expert_consensus"] * expert_sub
 + w["fundamental"] * fundamental_sub
 + w["technical"] * technical_sub
 + w["institutional"] * institutional_sub
 )
 quality = compute_data_quality_multiplier(confidence_grades or {})
 return min(max(raw * quality * 100, 0.0), 100.0)

Data Quality Multiplier

Each expert is assigned a data reliability grade (A/B/C/D), and the average grade is applied as a multiplier. When data quality is low, the composite score is automatically discounted.

Grade	Multiplier
A	1.00
B	0.85
C	0.60
D	0.30

4. UI Sliders and DB Migration

Weight Adjustment UI

Five per-factor weight sliders were added to the Settings page. As users move the sliders, the normalized proportions are displayed in real time. The existing min_rr_score slider was replaced with min_composite_score, with a default threshold of 15%.

Full-Stack Migration

Changing from min_rr_score to min_composite_score required modifications across all of the following layers.

Layer	File	Changes
Scoring module	`composite_score.py`	5 sub-scores + aggregation functions (new)
Scanner	`market_scanner.py`	Remove `compute_confidence`, connect composite score
Risk manager	`risk_manager.py`	Change gate criteria
API router	`agents.py`	Add weight fields, rename fields
Frontend types	`types.ts`	Add 5 weight fields
Settings UI	`SettingsView.tsx`	Add 5 sliders
DB	`trading.db`	Column rename + weight default inserts

5. Other Improvements

Alpha Pulse Rebranding

The project was renamed from “KIS Trading” to Alpha Pulse. All branding assets including favicon, manifest, header bar, and app title were replaced.

Infrastructure Fixes

APScheduler cron day-of-week conversion: Fixed schedule tasks to run on the correct day by converting between standard cron (0=Sun) and APScheduler (0=Mon) day-of-week indices
uvicorn WebSocket: Switched from websockets to wsproto implementation to resolve DeprecationWarning
Schedule sorting: Sorted schedule task list in ascending order by cron time (hour:minute)

Expert Panel Enhancements

Analysis quality was improved by providing each expert with additional data including investor flow trends, DART disclosure summaries, and specialty-specific confidence grades.

Commit Log

Date	Description	Category
03-24	Sort schedule tasks by cron time	UI
03-25	Make agent settings configurable + signal card UI improvements	feat
03-25	Update CLAUDE.md multi-agent system documentation	docs
03-30	Switch uvicorn WebSocket to wsproto	fix
03-30	Fix APScheduler cron day-of-week conversion	fix
03-30	Stock Info page design doc + implementation plan	docs
03-31	Add technical_service module (reusable indicator calculations)	feat
03-31	Add research types + API functions	feat
03-31	/api/research router (5 endpoints)	feat
03-31	9 stockinfo section components + DiscoverySidebar	feat
03-31	InsiderSection, SignalHistorySection components	feat
03-31	ResearchPanel, StockInfoView, CSS completion	feat
03-31	Connect StockInfoView to app navigation	feat
03-31	Resolve lint errors and complete stockinfo components	fix
03-31	Fix verbatimModuleSyntax compatible import type	fix
03-31	Return search results + prevent stale state on stock switch	fix
03-31	Signal pipeline fix design doc	docs
03-31	Candlestick chart + volume, MA, BB overlays	feat
03-31	Separate technical indicator mini chart cards	feat
03-31	Add compute_confidence linear mapping function	feat
03-31	Replace sigmoid with linear confidence, min_rr_score 0.3	feat
03-31	Technical indicator cards 2x2 grid layout	feat
03-31	SELL validation — reject unheld stocks + minimum hold time	feat
03-31	Force-convert unheld stock SELL to HOLD	feat
03-31	Add SELL/HOLD direction rules to Chief Analyst prompt	feat
03-31	Enhance expert data — flow, DART, confidence grades	feat
03-31	Adjust price/volume chart area spacing	fix
03-31	Calibration ceiling slider, min hold time input	feat
03-31	Fix missing RSI gauge CSS	fix
03-31	Multi-factor composite score design doc (Approach C)	docs
03-31	Multi-factor composite score implementation plan	docs
03-31	Rebrand from KIS Trading to Alpha Pulse	feat
04-01	5 sub-score functions + data quality multiplier	feat
04-01	compute_composite_score + weight normalization	feat
04-01	Connect composite score to pipeline, remove compute_confidence	feat
04-01	Change min_rr_score gate to min_composite_score (15%)	feat
04-01	Add weight fields to API router	feat
04-01	Add weight fields to frontend types	feat
04-01	Weight slider UI, replace min_composite_score	feat
04-01	DB migration — column rename + weight defaults	feat

Insights

Limitations of a single metric: Filtering trade signals with min_rr_score alone makes it impossible to distinguish stocks with high R/R but weak fundamentals, or stocks with good institutional flow but negative technical indicators. Transitioning to a multi-factor system allows evaluating each dimension independently and combining them via weighted sum. Users can adjust weights through sliders to tune according to their investment style (fundamental-focused vs. momentum-focused).

The value of reflecting data quality in scores: Not all factor data is of equal quality. For stocks with outdated DART disclosures, or stocks with low volume where technical indicators are unstable, a high composite score may not actually be reliable. Introducing a data quality multiplier to distinguish between “70 points calculated from good data” and “70 points calculated from poor data” was the core of this design.

The cost of field name changes that cut across the entire stack: Renaming a single min_rr_score to min_composite_score required modifications across 7 layers — DB, backend models, API router, frontend types, and UI components. Using more generic naming in the initial design could have reduced this cost.

stock-info on ICE-ICE-BEAR-BLOG