Chinese version

Theme scoring methodology

Overview

The scoring system reads five dimensions from ETL output raw_market.json and combines them into a weighted theme score from 0 to 100.

Dimensions

Dimension	Default weight	Source	Calculation
fundamental	0.30	Theme market performance	_scale(perf, midpoint=1.0, sensitivity=40)
valuation	0.25	Average P/E	_inverse_ratio_score(pe, baseline=35, sensitivity=90)
sentiment	0.20	Cboe Put/Call and AAII	Z-score, tanh compression, then mapping to 0–100
liquidity	0.15	Average P/S	_inverse_ratio_score(ps, baseline=8, sensitivity=70)
event	0.10	Configured constants	Fixed constants: AI 70, Magnificent 7 68, BTC 65

Scoring functions

_scale(value, midpoint, sensitivity)

Linear mapping to 0–100:

score = 50 + (value - midpoint) * sensitivity

The result is clamped to [0, 100].

_inverse_ratio_score(value, baseline, sensitivity)

For valuation metrics such as P/E and P/S, lower values score better:

_scale(baseline / value, midpoint=1.0, sensitivity)

Sentiment aggregation (sentiment.aggregate)

1. Put/Call ratio: calculate a z-score from the log ratio, invert it (panic is contrarian bullish), then apply tanh compression.
2. AAII bull-bear spread: calculate a z-score and invert it (extreme optimism is contrarian bearish), then apply tanh compression.
3. Combine components as 50 + 50 * mean(component_scores) and clamp to [0, 100].

History retains up to 252 Put/Call observations and 104 AAII observations.

Weight configuration

config/weights.yml supports per-theme overrides of the defaults:

weights:
  default:
    fundamental: 0.30
    valuation: 0.25
    sentiment: 0.20
    liquidity: 0.15
    event: 0.10
  theme_ai:
    fundamental: 0.30
    valuation: 0.15
    sentiment: 0.25
    liquidity: 0.20
    event: 0.10
  theme_btc:
    fundamental: 0.10
    valuation: 0.15
    sentiment: 0.30
    liquidity: 0.30
    event: 0.15

Thresholds and actions

Key	Default	Meaning
action_add	75	A total score at or above this value raises the attention level
action_trim	45	A total score at or below this value lowers the attention level

Fallback behavior

When ETL fails or data is missing:

• Missing dimensions use the neutral value 50 or a historical cache.
• scores.json.degraded is set to true.
• The daily report marks delayed data in its title and cards.
• A dimension with fallback: true indicates that a fallback value was used.

Updating the configuration

When changing config/weights.yml:

1. Update version and changed_at.
2. Run pytest -k contract to check contract tests.
3. Update expected values in tests/test_scoring.py.

Actual meanings of compatibility keys

These are descriptive heuristics, not out-of-sample validated predictions. AI fundamental measures market performance; Magnificent 7 measures aggregate market cap. Both equity liquidity keys measure P/S valuation, not trading liquidity or fund flow. BTC uses ETF net inflow. Event scores are fixed constants (AI 70, Magnificent 7 68, BTC 65), not measured event intensity. breakdown_detail includes the actual metric, display label, evidence level and limitations. See Agent access and score semantics.