fin-glassbox

Qualitative Synthesis Layer Analyst

1. Module Identity

File: code/analysts/qualitative_analyst.py
Module name: Qualitative Analyst
Project: An Explainable Multimodal Neural Framework for Financial Risk Management
Layer: Qualitative synthesis branch
Status: Trainable downstream analyst module
Consumes: Sentiment Analyst outputs + News Analyst outputs
Produces: Event-level and daily ticker-level qualitative signals for the Fusion Engine
Fundamentals used: No

The Qualitative Analyst is the trained synthesis module that combines text-side information from the Sentiment Analyst and News Analyst. It does not perform raw text encoding and it does not read SEC filings directly. Raw financial text is already encoded upstream by FinBERT, then processed by specialist downstream text analysts. This module learns how to combine those specialist outputs into a structured qualitative branch signal.

The module is intentionally separated from the Quantitative Analyst. The Quantitative Analyst uses technical, risk, and position-sizing outputs. The Qualitative Analyst uses text-derived outputs only.

2. Purpose in the Overall Architecture

The final system follows a modular, distributed design rather than a single monolithic model. The Qualitative Analyst exists to summarise textual market evidence into a form that can be merged later with the quantitative branch.

Its role is to answer:

“What does the available financial text evidence suggest about this ticker/date, and how risky or reliable is that text evidence?”

It does not answer:

“Should the final system buy or sell?”

That final action is handled later by the Fusion Engine and Final Trade Approver.

The simplified location of this module is:

FinBERT Encoder
    ↓
Sentiment Analyst        News Analyst
    ↓                    ↓
        Qualitative Analyst
                ↓
            Fusion Engine
                ↓
        Final Trade Approver

3. Why This Module Is Trainable

The first design possibility was a purely rule-based qualitative synthesis layer. However, training is important in the project architecture because each major analyst module should learn a calibrated mapping rather than only applying fixed rules.

The implemented design therefore uses a lightweight trainable MLP calibration model.

The model learns from structured upstream analyst outputs, not from raw text. This is important because the raw language understanding is already handled by FinBERT and the text-specific analysts.

The Qualitative Analyst is trained using weak-supervised targets derived from sentiment and news outputs. This makes the module trainable even when no manually labelled qualitative synthesis dataset exists.

This is defensible because:

  1. The module is not inventing labels from nothing; it learns from upstream analyst signals.
  2. The weak labels are transparent and formula-defined.
  3. The learned model can smooth, calibrate, and combine noisy upstream signals.
  4. The XAI output still exposes the original drivers and the learned feature importance.

4. Inputs

The module consumes two upstream prediction families.

4.1 Sentiment Analyst Outputs

Expected locations include:

outputs/results/analysts/sentiment/
outputs/results/SentimentAnalyst/

The script automatically searches for chunk/split-specific sentiment prediction files.

Expected file naming patterns include:

chunk{chunk}_{split}_predictions.csv
sentiment_predictions_chunk{chunk}_{split}.csv
predictions_chunk{chunk}_{split}.csv

Typical sentiment columns include:

ticker
filing_date or date
doc_id
accession
form_type
source_name
predicted_sentiment_score
predicted_sentiment_class
predicted_sentiment_confidence
predicted_sentiment_uncertainty
predicted_sentiment_magnitude

The module supports flexible column aliases. For example, predicted_sentiment_score, sentiment_score, and pred_sentiment_score can all be mapped to the internal sentiment_score field.

4.2 News Analyst Outputs

Expected locations include:

outputs/results/analysts/news/
outputs/results/NewsAnalyst/

Expected file naming patterns include:

chunk{chunk}_{split}_news_predictions.csv
news_predictions_chunk{chunk}_{split}.csv
predictions_chunk{chunk}_{split}.csv

Typical news columns include:

ticker
filing_date or date
doc_id
accession
form_type
source_name
predicted_news_event_impact
predicted_news_importance
predicted_risk_relevance
predicted_volatility_spike
predicted_drawdown_risk
predicted_news_uncertainty

The News Analyst is treated as an event-impact and event-risk source, not only a direction source.

5. Standardised Internal Feature Schema

The module converts all available sentiment and news columns into a fixed feature vector.

sentiment_score
sentiment_confidence
sentiment_uncertainty
sentiment_magnitude
sentiment_event_present
news_event_impact
news_importance
risk_relevance
volatility_spike
drawdown_risk
news_uncertainty
news_event_present
has_both_sources

These 13 features form the model input vector.

Feature Meanings

Feature Meaning
sentiment_score Directional sentiment score, normally from -1 to +1
sentiment_confidence Confidence of the Sentiment Analyst prediction
sentiment_uncertainty Uncertainty of the sentiment output
sentiment_magnitude Strength/magnitude of the sentiment signal
sentiment_event_present Binary indicator that sentiment evidence exists
news_event_impact Directional market impact estimated by News Analyst
news_importance Importance or relevance of the news/filing event
risk_relevance How relevant the news event is to financial risk
volatility_spike Estimated likelihood or severity of volatility increase
drawdown_risk Estimated downside/drawdown risk from the news event
news_uncertainty Uncertainty of the News Analyst output
news_event_present Binary indicator that news evidence exists
has_both_sources Binary indicator that both sentiment and news evidence exist

6. Output Targets

The model predicts three continuous outputs:

qualitative_score
qualitative_risk_score
qualitative_confidence

6.1 qualitative_score

Range:

-1 to +1

Interpretation:

-1 = strongly negative qualitative evidence
 0 = neutral or no clear qualitative direction
+1 = strongly positive qualitative evidence

6.2 qualitative_risk_score

Range:

0 to 1

Interpretation:

0 = low text-driven risk
1 = severe text-driven risk

This is separate from direction. A text event can be positive but still risky.

6.3 qualitative_confidence

Range:

0 to 1

Interpretation:

0 = unreliable or missing text evidence
1 = strong, confident qualitative evidence

7. Weak-Supervised Target Construction

Because there is no manually labelled qualitative synthesis dataset, the module constructs transparent weak-supervised targets from the upstream analyst outputs.

7.1 Direction Target

sentiment_direction = sentiment_score
news_direction = news_event_impact × news_importance
qualitative_score_target = 0.55 × sentiment_direction + 0.45 × news_direction

The result is clipped to:

[-1, +1]

Sentiment receives slightly higher weight because it is already a direct polarity signal, while news impact is event-based and may be noisier.

7.2 Risk Target

qualitative_risk_target =
    0.35 × risk_relevance
  + 0.30 × volatility_spike
  + 0.25 × drawdown_risk
  + 0.10 × news_uncertainty

The result is clipped to:

[0, 1]

This makes the model sensitive to text that warns of tail risk, instability, volatility, or downside pressure.

7.3 Confidence Target

qualitative_confidence_target =
    0.40 × sentiment_confidence
  + 0.25 × (1 - sentiment_uncertainty)
  + 0.20 × news_importance
  + 0.15 × (1 - news_uncertainty)

The result is clipped to:

[0, 1]

If neither sentiment nor news evidence exists, confidence is forced to 0.

8. Model Architecture

The Qualitative Analyst uses a small MLP.

Input: 13-dimensional feature vector
    ↓
Linear layer(s)
    ↓
Tanh activation
    ↓
Dropout
    ↓
Linear output head
    ↓
Output transformations

The output transformations are:

qualitative_score      = tanh(raw_output_0)
qualitative_risk       = sigmoid(raw_output_1)
qualitative_confidence = sigmoid(raw_output_2)

This ensures valid output ranges:

qualitative_score      ∈ [-1, +1]
qualitative_risk       ∈ [0, 1]
qualitative_confidence ∈ [0, 1]

Default architecture:

input_dim = 13
hidden_dim = 64
n_layers = 2
dropout = 0.15
activation = tanh
output_dim = 3

The hidden activation is tanh, consistent with the project preference for MLP hidden layers.

9. Training Objective

The module uses a weighted multi-output regression loss.

score_loss = MSE(predicted_qualitative_score, qualitative_score_target)
risk_loss  = MSE(predicted_qualitative_risk, qualitative_risk_target)
conf_loss  = MSE(predicted_qualitative_confidence, qualitative_confidence_target)

Final loss:

loss = 0.40 × score_loss + 0.35 × risk_loss + 0.25 × confidence_loss

The score and risk outputs receive higher emphasis because they are more directly used in the Fusion Engine.

10. Hyperparameter Optimisation

The module supports Optuna TPE hyperparameter search.

Search parameters:

hidden_dim:     [32, 64, 96, 128]
n_layers:       1 to 3
dropout:        0.05 to 0.40
learning_rate:  1e-4 to 3e-3
weight_decay:   1e-7 to 1e-3
batch_size:     [256, 512, 1024]

HPO is stored in SQLite:

outputs/codeResults/QualitativeAnalyst/hpo_chunk{chunk}.db

Best parameters are saved as:

outputs/codeResults/QualitativeAnalyst/best_params_chunk{chunk}.json

The HPO command is resumable unless --fresh is used.

11. Training Outputs

For each chunk, the module saves:

outputs/models/QualitativeAnalyst/chunk{chunk}/best_model.pt
outputs/models/QualitativeAnalyst/chunk{chunk}/final_model.pt
outputs/models/QualitativeAnalyst/chunk{chunk}/scaler.npz
outputs/models/QualitativeAnalyst/chunk{chunk}/training_history.csv
outputs/models/QualitativeAnalyst/chunk{chunk}/model_freezed/model.pt
outputs/models/QualitativeAnalyst/chunk{chunk}/model_unfreezed/model.pt

The model checkpoint contains:

state_dict
config
feature_names
target_names
best_val_loss

The scaler contains train-fitted feature mean and standard deviation. The scaler is fitted only on the training split to avoid validation/test leakage.

12. Prediction Outputs

The module produces two prediction outputs per chunk/split.

12.1 Event-Level Output

Path:

outputs/results/QualitativeAnalyst/qualitative_events_chunk{chunk}_{split}.csv

Purpose:

Preserve individual text-event evidence for explanation and audit.

Important columns:

ticker
date
chunk
split
doc_id
accession
form_type
source_name
sentiment_score
sentiment_confidence
sentiment_uncertainty
sentiment_magnitude
news_event_impact
news_importance
risk_relevance
volatility_spike
drawdown_risk
news_uncertainty
sentiment_direction_score
news_direction_score
qualitative_score
qualitative_risk_score
qualitative_confidence
qualitative_recommendation
dominant_qualitative_driver
xai_summary

12.2 Daily Ticker-Level Output

Path:

outputs/results/QualitativeAnalyst/qualitative_daily_chunk{chunk}_{split}.csv

Purpose:

Provide ticker/date aligned qualitative signals for the Fusion Engine.

Important columns:

ticker
date
chunk
split
event_count
sentiment_event_count
news_event_count
qualitative_score
qualitative_risk_score
qualitative_confidence
qualitative_recommendation
max_event_risk_score
mean_event_risk_score
mean_sentiment_score
mean_news_impact_score
mean_news_importance
dominant_qualitative_driver
xai_summary

13. Daily Aggregation Logic

The event-level output is aggregated by:

ticker, date

13.1 Daily Qualitative Score

daily_qualitative_score = confidence-weighted mean of event qualitative_score

13.2 Daily Qualitative Risk

daily_qualitative_risk_score =
    0.60 × max(event_qualitative_risk_score)
  + 0.40 × weighted_mean(event_qualitative_risk_score)

This prevents severe risk disclosures from being averaged away by many neutral events.

13.3 Daily Qualitative Confidence

daily_qualitative_confidence = mean(event_confidence) × min(1, log1p(event_count) / log(5))

This rewards repeated supporting evidence but caps the boost.

14. Recommendation Logic

The module converts qualitative score, risk, and confidence into a qualitative recommendation.

BUY:
    qualitative_score > +0.20
    qualitative_risk_score < 0.70
    qualitative_confidence ≥ 0.40

SELL:
    qualitative_score < -0.20
    OR qualitative_risk_score > 0.85

HOLD:
    otherwise

This conservative logic ensures that the qualitative branch does not overreact to weak or uncertain text signals.

15. Explainability Design

The module provides both row-level and file-level XAI.

15.1 Row-Level XAI

Each event and daily row contains an xai_summary column.

Example:

SELL: qualitative_score=-0.421, risk=0.881, confidence=0.624; driver=drawdown_risk; sentiment=-0.350; news_impact=-0.220.

15.2 Dominant Driver

Each event row includes:

dominant_qualitative_driver

Possible values include:

sentiment
news_impact
risk_relevance
volatility_spike
drawdown_risk
news_uncertainty

This tells which upstream factor most influenced the qualitative output.

15.3 Gradient-Based Feature Importance

The module computes input-gradient feature importance for a sample of rows.

XAI output path:

outputs/results/QualitativeAnalyst/xai/qualitative_chunk{chunk}_{split}_xai_summary.json

The JSON contains:

module
chunk
split
config
model metadata
event row count
daily row count
ticker count
date range
event recommendation counts
daily recommendation counts
dominant driver counts
gradient feature importance
top positive event examples
top negative event examples
highest risk event examples
plain-English explanation
summary statistics

16. Integration with Fusion Engine

The Fusion Engine should consume the daily qualitative file:

outputs/results/QualitativeAnalyst/qualitative_daily_chunk{chunk}_{split}.csv

It should merge on:

ticker, date

with:

outputs/results/QuantitativeAnalyst/quantitative_analysis_chunk{chunk}_{split}.csv
outputs/results/PositionSizing/position_sizing_chunk{chunk}_{split}.csv

Important interpretation rule:

Missing qualitative row = no text evidence available

It should not be interpreted as positive or negative sentiment.

Recommended Fusion defaults for missing qualitative evidence:

qualitative_score = 0.0
qualitative_confidence = 0.0
qualitative_risk_score = 0.5 or neutral default

This ensures that absent text does not falsely strengthen the decision.

17. No Fundamentals Policy

This module intentionally excludes fundamentals.

It does not read:

fundamental embeddings
fundamental analyst outputs
SEC tabular fundamentals
valuation ratios
balance-sheet strength outputs

The current qualitative branch is limited to:

Sentiment Analyst + News Analyst

This preserves the current approved architecture.

18. Leakage Prevention

The module follows these anti-leakage rules:

  1. It trains on the training split and validates on the validation split.
  2. The feature scaler is fitted only on training data.
  3. Validation and test features are transformed using the saved training scaler.
  4. The module uses existing upstream predictions rather than future market outcomes.
  5. It outputs event-level and daily-level signals using the event/filing date only.

The weak-supervised targets are derived from same-row upstream analyst outputs, so they do not introduce future price information.

19. Strengths

The Qualitative Analyst has several strengths:

  1. It is trainable, satisfying the architecture requirement for a learned analyst layer.
  2. It is lightweight and fast compared to retraining FinBERT.
  3. It remains explainable because all inputs are interpretable upstream analyst scores.
  4. It produces both event-level and daily-level outputs.
  5. It is fusion-ready because daily outputs are ticker/date aligned.
  6. It supports HPO, checkpointing, scaler persistence, validation, and XAI.

20. Limitations

This version has limitations that should be acknowledged in the report.

  1. It uses weak-supervised targets, not manually labelled qualitative decisions.
  2. It depends on the quality of upstream Sentiment and News Analyst outputs.
  3. It does not directly read article text or SEC filing text.
  4. It does not include fundamentals.
  5. Text coverage is sparse, so many ticker-days may not have qualitative evidence.
  6. If sentiment/news models are rerun, this module should also be rerun.

These limitations are acceptable because this module is a synthesis component, not a primary text encoder.

21. Main CLI Commands

Compile

cd ~/fin-glassbox && python -m py_compile code/analysts/qualitative_analyst.py

Inspect Inputs

cd ~/fin-glassbox && python code/analysts/qualitative_analyst.py inspect --repo-root .

Smoke Test

cd ~/fin-glassbox && python code/analysts/qualitative_analyst.py smoke --repo-root . --device cuda

HPO for One Chunk

cd ~/fin-glassbox && python code/analysts/qualitative_analyst.py hpo --repo-root . --chunk 1 --trials 30 --device cuda --fresh

Train One Chunk with Best HPO Parameters

cd ~/fin-glassbox && python code/analysts/qualitative_analyst.py train-best --repo-root . --chunk 1 --device cuda --fresh

Predict One Split

cd ~/fin-glassbox && python code/analysts/qualitative_analyst.py predict --repo-root . --chunk 1 --split test --device cuda

Validate One Output

cd ~/fin-glassbox && python code/analysts/qualitative_analyst.py validate --repo-root . --chunk 1 --split test

Before launching the full run, execute a small sanity run:

cd ~/fin-glassbox && python -m py_compile code/analysts/qualitative_analyst.py && python code/analysts/qualitative_analyst.py inspect --repo-root . && python code/analysts/qualitative_analyst.py smoke --repo-root . --device cuda

Then run one chunk with a small number of trials:

cd ~/fin-glassbox && python code/analysts/qualitative_analyst.py hpo --repo-root . --chunk 1 --trials 3 --device cuda --fresh && python code/analysts/qualitative_analyst.py train-best --repo-root . --chunk 1 --device cuda --fresh && python code/analysts/qualitative_analyst.py predict --repo-root . --chunk 1 --split val --device cuda && python code/analysts/qualitative_analyst.py predict --repo-root . --chunk 1 --split test --device cuda && python code/analysts/qualitative_analyst.py validate --repo-root . --chunk 1 --split test

If this passes, run the full version.

23. Full Qualitative Analyst Run

Use this after sentiment and news outputs exist for all chunks.

cd ~/fin-glassbox && python code/analysts/qualitative_analyst.py hpo --repo-root . --chunk 1 --trials 30 --device cuda --fresh && python code/analysts/qualitative_analyst.py train-best --repo-root . --chunk 1 --device cuda --fresh && python code/analysts/qualitative_analyst.py predict-all --repo-root . --chunks 1 --splits val test --device cuda && python code/analysts/qualitative_analyst.py hpo --repo-root . --chunk 2 --trials 30 --device cuda --fresh && python code/analysts/qualitative_analyst.py train-best --repo-root . --chunk 2 --device cuda --fresh && python code/analysts/qualitative_analyst.py predict-all --repo-root . --chunks 2 --splits val test --device cuda && python code/analysts/qualitative_analyst.py hpo --repo-root . --chunk 3 --trials 30 --device cuda --fresh && python code/analysts/qualitative_analyst.py train-best --repo-root . --chunk 3 --device cuda --fresh && python code/analysts/qualitative_analyst.py predict-all --repo-root . --chunks 3 --splits val test --device cuda

24. Comprehensive Upstream Rerun Command

After FinBERT, Sentiment Analyst, and News Analyst are fully rerun, the complete qualitative branch can be refreshed with this command:

cd ~/fin-glassbox && python code/analysts/sentiment_analyst.py hpo --repo-root . --chunk 1 --trials 30 --device cuda --fresh && python code/analysts/sentiment_analyst.py train-best --repo-root . --chunk 1 --device cuda --fresh && python code/analysts/sentiment_analyst.py predict --repo-root . --chunk 1 --split val --device cuda && python code/analysts/sentiment_analyst.py predict --repo-root . --chunk 1 --split test --device cuda && python code/analysts/news_analyst.py hpo --repo-root . --chunk 1 --trials 30 --device cuda --fresh && python code/analysts/news_analyst.py train-best --repo-root . --chunk 1 --device cuda --fresh && python code/analysts/news_analyst.py predict --repo-root . --chunk 1 --split val --device cuda && python code/analysts/news_analyst.py predict --repo-root . --chunk 1 --split test --device cuda && python code/analysts/qualitative_analyst.py hpo --repo-root . --chunk 1 --trials 30 --device cuda --fresh && python code/analysts/qualitative_analyst.py train-best --repo-root . --chunk 1 --device cuda --fresh && python code/analysts/qualitative_analyst.py predict-all --repo-root . --chunks 1 --splits val test --device cuda && python code/analysts/sentiment_analyst.py hpo --repo-root . --chunk 2 --trials 30 --device cuda --fresh && python code/analysts/sentiment_analyst.py train-best --repo-root . --chunk 2 --device cuda --fresh && python code/analysts/sentiment_analyst.py predict --repo-root . --chunk 2 --split val --device cuda && python code/analysts/sentiment_analyst.py predict --repo-root . --chunk 2 --split test --device cuda && python code/analysts/news_analyst.py hpo --repo-root . --chunk 2 --trials 30 --device cuda --fresh && python code/analysts/news_analyst.py train-best --repo-root . --chunk 2 --device cuda --fresh && python code/analysts/news_analyst.py predict --repo-root . --chunk 2 --split val --device cuda && python code/analysts/news_analyst.py predict --repo-root . --chunk 2 --split test --device cuda && python code/analysts/qualitative_analyst.py hpo --repo-root . --chunk 2 --trials 30 --device cuda --fresh && python code/analysts/qualitative_analyst.py train-best --repo-root . --chunk 2 --device cuda --fresh && python code/analysts/qualitative_analyst.py predict-all --repo-root . --chunks 2 --splits val test --device cuda && python code/analysts/sentiment_analyst.py hpo --repo-root . --chunk 3 --trials 30 --device cuda --fresh && python code/analysts/sentiment_analyst.py train-best --repo-root . --chunk 3 --device cuda --fresh && python code/analysts/sentiment_analyst.py predict --repo-root . --chunk 3 --split val --device cuda && python code/analysts/sentiment_analyst.py predict --repo-root . --chunk 3 --split test --device cuda && python code/analysts/news_analyst.py hpo --repo-root . --chunk 3 --trials 30 --device cuda --fresh && python code/analysts/news_analyst.py train-best --repo-root . --chunk 3 --device cuda --fresh && python code/analysts/news_analyst.py predict --repo-root . --chunk 3 --split val --device cuda && python code/analysts/news_analyst.py predict --repo-root . --chunk 3 --split test --device cuda && python code/analysts/qualitative_analyst.py hpo --repo-root . --chunk 3 --trials 30 --device cuda --fresh && python code/analysts/qualitative_analyst.py train-best --repo-root . --chunk 3 --device cuda --fresh && python code/analysts/qualitative_analyst.py predict-all --repo-root . --chunks 3 --splits val test --device cuda

25. Expected Final Outputs

After a successful full run, the following files should exist:

outputs/models/QualitativeAnalyst/chunk1/best_model.pt
outputs/models/QualitativeAnalyst/chunk2/best_model.pt
outputs/models/QualitativeAnalyst/chunk3/best_model.pt

outputs/results/QualitativeAnalyst/qualitative_events_chunk1_val.csv
outputs/results/QualitativeAnalyst/qualitative_events_chunk1_test.csv
outputs/results/QualitativeAnalyst/qualitative_events_chunk2_val.csv
outputs/results/QualitativeAnalyst/qualitative_events_chunk2_test.csv
outputs/results/QualitativeAnalyst/qualitative_events_chunk3_val.csv
outputs/results/QualitativeAnalyst/qualitative_events_chunk3_test.csv

outputs/results/QualitativeAnalyst/qualitative_daily_chunk1_val.csv
outputs/results/QualitativeAnalyst/qualitative_daily_chunk1_test.csv
outputs/results/QualitativeAnalyst/qualitative_daily_chunk2_val.csv
outputs/results/QualitativeAnalyst/qualitative_daily_chunk2_test.csv
outputs/results/QualitativeAnalyst/qualitative_daily_chunk3_val.csv
outputs/results/QualitativeAnalyst/qualitative_daily_chunk3_test.csv

outputs/results/QualitativeAnalyst/xai/qualitative_chunk1_val_xai_summary.json
outputs/results/QualitativeAnalyst/xai/qualitative_chunk1_test_xai_summary.json
outputs/results/QualitativeAnalyst/xai/qualitative_chunk2_val_xai_summary.json
outputs/results/QualitativeAnalyst/xai/qualitative_chunk2_test_xai_summary.json
outputs/results/QualitativeAnalyst/xai/qualitative_chunk3_val_xai_summary.json
outputs/results/QualitativeAnalyst/xai/qualitative_chunk3_test_xai_summary.json

26. Validation Checklist

A qualitative run is considered successful if:

1. Training completes and best_model.pt is saved.
2. scaler.npz exists for each chunk.
3. Event-level outputs exist for validation and test splits.
4. Daily outputs exist for validation and test splits.
5. XAI JSON outputs exist for validation and test splits.
6. qualitative_score is always within [-1, +1].
7. qualitative_risk_score is always within [0, 1].
8. qualitative_confidence is always within [0, 1].
9. Recommendation labels are only BUY, HOLD, SELL.
10. Fusion can merge daily qualitative outputs on ticker/date.

27. Report Description

The Qualitative Analyst is a trainable synthesis module that combines outputs from the Sentiment Analyst and News Analyst into a daily ticker-level qualitative signal. Rather than directly reprocessing raw text, it learns a calibrated mapping from upstream text-specialist predictions into qualitative direction, qualitative risk, and qualitative confidence. The module is trained using transparent weak-supervised targets derived from sentiment polarity, news impact, risk relevance, volatility-spike likelihood, drawdown-risk likelihood, and uncertainty. It produces event-level outputs for auditability and ticker-date daily outputs for integration into the Fusion Engine. Explainability is provided through row-level explanations, dominant-driver attribution, and gradient-based feature importance.

28. Academic Talking Points

  1. The module separates qualitative text evidence from quantitative market/risk evidence.
  2. It does not duplicate FinBERT; it uses the specialised outputs already produced by FinBERT-based analysts.
  3. It is trainable, but still interpretable because its input features are human-readable analyst scores.
  4. It outputs direction, risk, and confidence separately.
  5. Daily aggregation prevents the Fusion Engine from needing to handle raw event-level text rows.
  6. XAI is built in through dominant-driver explanations and gradient feature importance.
  7. Missing qualitative evidence is treated as absence of text evidence, not as positive or negative information.
  8. The module can be rerun easily whenever FinBERT, Sentiment Analyst, or News Analyst outputs improve.

29. Summary

The Qualitative Analyst completes the qualitative branch of the architecture. It transforms sparse, event-level text analysis into structured, trainable, explainable daily qualitative signals. It is designed to be lightweight, reproducible, auditable, and directly compatible with the upcoming Fusion Engine.

Its main contribution is:

Sentiment + News → trained qualitative direction/risk/confidence → Fusion-ready daily signal

This keeps the full system modular, explainable, and defensible.

This site is open source. Improve this page.