docs/ta_reference/institutional_hotspot_rules.md

Institutional Hotspot Rules

Date: 2026-05-21

Hierarchy entry point: docs/hotspot/README.md

Purpose

Translate 120-day Taiwanese institutional-flow behavior into testable hotspot scanner features. These rules are hypotheses for ranking and validation; they are not production trading advice.

Data Contract

Each rule is evaluated as of a completed trading day.

Required input columns:

• foreign_net
• trust_net
• dealer_net
• institutional_net
• close
• ma10
• ma20
• volume
• turnover

Rolling windows:

short flow = 5 trading days
medium flow = 20 trading days
long flow = 60 trading days
institutional memory = 120 trading days

Missing flow data should produce neutral features and an explicit skipped or partial-data reason in validation output.

Accumulation Rules

120-Day Combined Accumulation

flow_120d_accumulation =
  institutional_net_120d > 0
  and institutional_net_120d_percentile >= 70

Interpretation:

• Long-window institutional demand exists.
• Use as a rank boost only when liquidity coverage is acceptable.

20-Day Acceleration

flow_20d_acceleration =
  institutional_net_20d > 0
  and institutional_net_20d / 20 > institutional_net_120d / 120

Interpretation:

• Recent buying is stronger than the long-window pace.
• Use with price/volume confirmation to avoid stale accumulation.

Institution-Specific Support Rules

Investment Trust 10MA Support

trust_10ma_support =
  trust_net_5d > 0
  and trust_net_20d > 0
  and close >= ma10

Risk variant:

trust_10ma_break =
  trust_net_20d > 0
  and close < ma10

Interpretation:

• Trust buying above 10MA is supportive.
• A 10MA break after trust buying is a hotspot penalty, not a short signal by itself.

Foreign 20MA Support

foreign_20ma_support =
  foreign_net_20d > 0
  and close >= ma20

Risk variant:

foreign_20ma_break =
  foreign_net_20d > 0
  and close < ma20

Interpretation:

• Foreign accumulation above 20MA is supportive.
• A 20MA break after foreign accumulation should lower rank confidence.

Dealer Confirmation

dealer_confirmation =
  dealer_net_5d > 0
  and (foreign_net_5d > 0 or trust_net_5d > 0)

Interpretation:

• Dealer buying confirms another institutional buyer.
• Dealer-only buying should be weaker than foreign or trust confirmation until validation proves otherwise.

Divergence And Risk Rules

Flow Divergence Risk

flow_divergence_risk =
  close_20d_return > 0
  and institutional_net_20d <= 0

Distribution Risk

flow_distribution_risk =
  close >= close_60d_high * 0.95
  and institutional_net_20d < institutional_net_60d / 3

Interpretation:

• Price near highs with fading institutional support is a rank penalty.
• Keep this separate from bearish prediction labels.

Volume-Spike Fade Breakdown

volume_spike_fade_breakdown =
  volume[t-3] >= 2.0 * average_volume[t-23:t-4]
  and volume[t-2] <= 0.75 * volume[t-3]
  and high[t-2] <= high[t-3]
  and close[t-1] < open[t-1]
  and close[t] < open[t]
  and close[t] < ma20[t]

Interpretation:

• The earlier volume spike did not attract follow-through buying.
• Two black candles plus MA20 break means the hotspot may have shifted away.
• Use as a hotspot rank penalty and display risk flag, not a hard sell rule.

Hotspot Score Composition

Initial experimental score:

hotspot_score =
  volume_turnover_rank
  + flow_120d_accumulation
  + flow_20d_acceleration
  + trust_10ma_support
  + foreign_20ma_support
  + dealer_confirmation
  - flow_divergence_risk
  - flow_distribution_risk
  - volume_spike_fade_breakdown

The exact weights must be learned or grid-searched in the harness. Do not hard-code permanent production weights without validation.

Validation Requirements

• Use the no-lookahead split from docs/full_market_hotspot_harness.md.
• Evaluate top 20, top 30, and top 50 hotspot lists.
• Compare against the current local hot-stock ranking baseline.
• Record per-rule ablations so a single noisy flow rule can be removed without discarding the whole scanner.
• Promotion requires the gates in docs/full_market_hotspot_harness.md.