data-science-for-intelligence

name	data-science-for-intelligence
description	Statistical analysis, ML, NLP, time series forecasting, network analysis for political intelligence data
license	Apache-2.0

Data Science for Intelligence Skill

🔴 AI FIRST Quality Principle

This skill MUST be applied with the AI FIRST principle: never accept first-pass quality. ALL analysis and content MUST go through minimum 2 complete iterations. After first pass, read ALL output back completely and systematically improve every section — strengthen evidence, deepen analysis, add specific citations, broaden perspectives. Spend ALL allocated time on real work. Single-pass output is NEVER acceptable. NO SHORTCUTS.

Purpose

This skill provides comprehensive data science methodologies tailored for political intelligence analysis in the Riksdagsmonitor platform. It covers statistical analysis, machine learning, natural language processing, time series forecasting, and network analysis techniques applied to the 6 intelligence frameworks and 82 database views for democratic accountability assessment.

When to Use This Skill

Apply this skill when:

• ✅ Building predictive models for election outcomes or coalition stability
• ✅ Analyzing temporal patterns in voting behavior or policy positions
• ✅ Detecting anomalies in political behavior (sudden voting shifts, absences)
• ✅ Performing text analysis on parliamentary documents and speeches
• ✅ Constructing influence networks from voting alignment patterns
• ✅ Forecasting trends in party support or politician effectiveness
• ✅ Clustering politicians or parties by voting similarity
• ✅ Modelling Swedish economic transmission chains using IMF macro context plus SCB KPI/AKU/HEK/fuel-price series and Riksbank policy-rate/minutes signals

Do NOT use for:

• ❌ Simple aggregation queries (use SQL views instead)
• ❌ Real-time operational dashboards (use materialized views)
• ❌ Causal inference without proper experimental design or quasi-experimental methods
• ❌ Replacing IMF cross-country macro/fiscal indicators with SCB or Riksbank series; use SCB/Riksbank only as Swedish ground-truth complements

Data Science Framework for CIA Platform

6 Intelligence Analysis Frameworks

graph TB
    subgraph "Data Sources"
        A[Riksdagen API<br/>3.5M votes]
        B[Election Authority<br/>40 parties]
        C[World Bank<br/>598K indicators]
        D[Financial Authority<br/>Agency data]
    end
    
    subgraph "Data Science Techniques"
        A & B & C & D --> E[Feature Engineering]
        E --> F1[Time Series<br/>Analysis]
        E --> F2[Classification<br/>Models]
        E --> F3[Clustering<br/>Analysis]
        E --> F4[NLP<br/>Processing]
        E --> F5[Network<br/>Analysis]
    end
    
    subgraph "Intelligence Frameworks"
        F1 --> G1[1. Temporal<br/>Analysis]
        F2 --> G2[3. Pattern<br/>Recognition]
        F3 --> G3[2. Comparative<br/>Analysis]
        F4 --> G3
        F5 --> G5[5. Network<br/>Analysis]
        F1 & F2 --> G4[4. Predictive<br/>Intelligence]
        G1 & G2 & G3 & G4 & G5 --> G6[6. Decision<br/>Intelligence]
    end
    
    subgraph "Intelligence Products"
        G1 & G2 & G3 & G4 & G5 & G6 --> H[Risk Assessments]
        H --> I[Political Scorecards]
        H --> J[Coalition Forecasts]
        H --> K[Anomaly Alerts]
    end
    
    style E fill:#ffeb99
    style F1 fill:#e1f5ff
    style F2 fill:#e1f5ff
    style F3 fill:#e1f5ff
    style F4 fill:#e1f5ff
    style F5 fill:#e1f5ff
    style H fill:#ccffcc

1. Time Series Analysis (Temporal Framework)

Purpose: Analyze trends, seasonality, and forecast political metrics over time.

CIA Platform Applications:

• Politician voting participation trends
• Party support trajectories
• Government approval ratings
• Legislative productivity patterns

Decomposition Analysis

Example: Decompose Party Support into Trend, Seasonal, Residual Components

import pandas as pd
import numpy as np
from statsmodels.tsa.seasonal import seasonal_decompose
from statsmodels.tsa.stattools import adfuller
import matplotlib.pyplot as plt

class PoliticalTimeSeriesAnalyzer:
    """
    Time series analysis for political intelligence
    Supports: Temporal Analysis Framework
    """
    
    def __init__(self, db_connection):
        self.db = db_connection
    
    def decompose_party_support(self, party_code, election_years):
        """
        Decompose historical party support into trend, seasonal, residual
        
        Data Source: sweden_political_party table
        Intelligence Framework: Temporal Analysis
        """
        
        # Query: Historical election results
        query = """
        SELECT 
            election_year,
            percentage as support_percentage
        FROM sweden_political_party
        WHERE party_name = (SELECT party_name FROM sweden_political_party WHERE party_id = %s LIMIT 1)
            AND election_year >= %s
        ORDER BY election_year
        """
        
        df = pd.read_sql(query, self.db, params=[party_code, min(election_years)])
        df['election_year'] = pd.to_datetime(df['election_year'], format='%Y')
        df.set_index('election_year', inplace=True)
        
        # Perform seasonal decomposition (additive model)
        decomposition = seasonal_decompose(
            df['support_percentage'], 
            model='additive', 
            period=3  
        )
        
         {
            : decomposition.trend,
            : decomposition.seasonal,
            : decomposition.resid,
            : df[]
        }
    
     ():
        
         statsmodels.tsa.arima.model  ARIMA
        
        
        query = 
        
        df = pd.read_sql(query, .db, params=[party_code])
        
        
        model = ARIMA(df[], order=(, , ))
        fitted_model = model.fit()
        
        
        forecast = fitted_model.forecast(steps=forecast_periods)
        confidence_interval = fitted_model.get_forecast(steps=forecast_periods).conf_int()
        
         {
            : forecast,
            : confidence_interval.iloc[:, ],
            : confidence_interval.iloc[:, ],
            : fitted_model.summary()
        }
    
     ():
        
         ruptures  Pelt
        
        
        query = 
        
        df = pd.read_sql(query, .db, params=[person_id])
        signal = df[].values
        
        
        algo = Pelt(model=).fit(signal)
        changepoints = algo.predict(pen=)
        
        
        changepoint_dates = [df.iloc[cp][]  cp  changepoints[:-]]
        
         {
            : changepoints,
            : changepoint_dates,
            : signal
        }

SQL Time Series Query Example:

-- Calculate 12-month rolling average of party voting success rate
WITH monthly_performance AS (
    SELECT 
        party,
        DATE_TRUNC('month', vote_date) as month,
        AVG(CASE WHEN won = TRUE THEN 1.0 ELSE 0.0 END) as win_rate
    FROM view_riksdagen_party_ballot_support_annual_summary
    WHERE vote_date >= CURRENT_DATE - INTERVAL '4 years'
    GROUP BY party, DATE_TRUNC('month', vote_date)
)
SELECT 
    party,
    month,
    win_rate,
    AVG(win_rate) OVER (
        PARTITION BY party 
        ORDER BY month 
        ROWS BETWEEN 11 PRECEDING AND CURRENT ROW
    ) as rolling_12m_avg,
    win_rate - AVG(win_rate) OVER (
        PARTITION BY party 
        ORDER BY month 
           PRECEDING   
    )  deviation_from_trend
 monthly_performance
  party,  ;

2. Classification Models (Pattern Recognition Framework)

Purpose: Classify politicians, parties, or votes into predefined categories based on features.

CIA Platform Applications:

• Predict MP defection risk (HIGH/MEDIUM/LOW)
• Classify voting behavior as LOYAL/INDEPENDENT/REBELLIOUS
• Identify party switcher candidates
• Detect at-risk coalition members

Random Forest Classifier

Example: Predict MP Defection Risk

from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split, cross_val_score
from sklearn.metrics import classification_report, confusion_matrix
import pandas as pd

class PoliticianRiskClassifier:
    """
    Machine learning classification for behavioral risk assessment
    Supports: Pattern Recognition Framework
    """
    
    def __init__(self, db_connection):
        self.db = db_connection
        self.model = None
    
    def prepare_defection_features(self):
        """
        Feature engineering for defection risk prediction
        
        Data Sources:
        - view_riksdagen_politician_summary
        - view_riksdagen_party_coalition_agreeableness
        - view_riksdagen_politician_document_daily_summary
        """
        
        query = """
        WITH politician_metrics AS (
            SELECT 
                p.person_id,
                p.first_name || ' ' || p.last_name as name,
                p.party,
                p.total_days_served,
                p.total_assignments,
                p.total_ballots,
                p.total_documents,
                p.percent_absent,
                p.percent_abstain,
                EXTRACT(YEAR FROM AGE(CURRENT_DATE, p.born)) as age
            FROM view_riksdagen_politician_summary p
            WHERE p.status = 'Tjänstgörande riksdagsledamot'
        ),
        party_cohesion AS (
            SELECT 
                party,
                AVG(party_avg_agreement) as cohesion_score
            FROM view_riksdagen_party_coalition_agreeableness
            GROUP BY party
        ),
        recent_activity AS (
            SELECT 
                person_id,
                AVG(total_document_activity) as avg_monthly_activity,
                STDDEV(total_document_activity) as activity_volatility
            FROM view_riksdagen_politician_document_daily_summary
            WHERE active_date >= CURRENT_DATE - INTERVAL '6 months'
            GROUP BY person_id
        )
        SELECT 
            pm.*,
            pc.cohesion_score as party_cohesion,
            ra.avg_monthly_activity,
            ra.activity_volatility,
            CASE 
                -- Label based on historical defections (ground truth)
                WHEN pm.party != LAG(pm.party) OVER (PARTITION BY pm.person_id ORDER BY pm.total_days_served) 
                    THEN 1 
                ELSE 0 
            END as defected_label
        FROM politician_metrics pm
        LEFT JOIN party_cohesion pc ON pm.party = pc.party
        LEFT JOIN recent_activity ra ON pm.person_id = ra.person_id
        """
        
        df = pd.read_sql(query, .db)
        
        
        features = [
            , , , 
            , , ,
            , , , 
        ]
        
        X = df[features].fillna()
        y = df[]
        
         X, y, df[[, , ]]
    
     ():
        
        X, y, metadata = .prepare_defection_features()
        
        
        X_train, X_test, y_train, y_test = train_test_split(
            X, y, test_size=, random_state=, stratify=y
        )
        
        
        .model = RandomForestClassifier(
            n_estimators=,
            max_depth=,
            min_samples_split=,
            class_weight=,  
            random_state=
        )
        
        .model.fit(X_train, y_train)
        
        
        y_pred = .model.predict(X_test)
        
        
        cv_scores = cross_val_score(.model, X, y, cv=, scoring=)
        
         {
            : .model.score(X_test, y_test),
            : classification_report(y_test, y_pred),
            : confusion_matrix(y_test, y_pred),
            : cv_scores.mean(),
            : cv_scores.std(),
            : ((X.columns, .model.feature_importances_))
        }
    
     ():
        
        X, y, metadata = .prepare_defection_features()
        
        
        probabilities = .model.predict_proba(X)[:, ]  
        
        
        risk_report = metadata.copy()
        risk_report[] = probabilities
        risk_report[] = pd.cut(
            probabilities,
            bins=[, , , ],
            labels=[, , ]
        )
        
         risk_report.sort_values(, ascending=)

3. Clustering Analysis (Comparative Framework)

Purpose: Group similar politicians or parties without predefined labels.

CIA Platform Applications:

• Identify ideological clusters beyond formal party affiliations
• Discover voting blocs within parties
• Group politicians by policy focus
• Detect coalition alignment patterns

K-Means Clustering

Example: Cluster Politicians by Voting Behavior

from sklearn.cluster import KMeans
from sklearn.preprocessing import StandardScaler
from sklearn.decomposition import PCA
import pandas as pd
import numpy as np

class PoliticalClusteringAnalyzer:
    """
    Unsupervised clustering for political alignment discovery
    Supports: Comparative Analysis Framework
    """
    
    def __init__(self, db_connection):
        self.db = db_connection
    
    def cluster_politicians_by_voting(self, n_clusters=8):
        """
        K-means clustering of politicians based on voting patterns
        
        Data Source: vote_data table
        Intelligence Application: Identify cross-party voting blocs
        """
        
        # Query: Build voting similarity matrix
        query = """
        WITH politician_vote_vectors AS (
            SELECT 
                person_id,
                ballot_id,
                CASE vote 
                    WHEN 'Ja' THEN 1
                    WHEN 'Nej' THEN -1
                    WHEN 'Avstår' THEN 0
                    ELSE NULL  -- Exclude absences
                END as vote_value
            FROM vote_data
            WHERE vote_date >= CURRENT_DATE - INTERVAL '12 months'
                AND vote IN ('Ja', 'Nej', 'Avstår')
        ),
        pivot_votes AS (
            SELECT 
                person_id,
                ballot_id,
                vote_value
            FROM politician_vote_vectors
        )
        SELECT 
            pv.person_id,
            p.first_name || ' ' || p.last_name as name,
            p.party,
            -- Aggregate vote patterns (one row per person)
            AVG(CASE WHEN pv.vote_value = 1 THEN 1.0 ELSE 0.0 END) as yes_rate,
            AVG(CASE WHEN pv.vote_value = -1 THEN 1.0 ELSE 0.0 END) as no_rate,
            AVG(CASE WHEN pv.vote_value = 0 THEN 1.0 ELSE 0.0 END) as abstain_rate,
            COUNT(DISTINCT pv.ballot_id) as ballots_participated
        FROM pivot_votes pv
        JOIN person_data p ON pv.person_id = p.person_id
        GROUP BY pv.person_id, p.first_name, p.last_name, p.party
        HAVING COUNT(DISTINCT pv.ballot_id) > 100  -- Active MPs only
        """
        
        df = pd.read_sql(query, .db)
        
        
        features = [, , , ]
        X = df[features]
        
        
        scaler = StandardScaler()
        X_scaled = scaler.fit_transform(X)
        
        
        kmeans = KMeans(n_clusters=n_clusters, random_state=, n_init=)
        df[] = kmeans.fit_predict(X_scaled)
        
        
        pca = PCA(n_components=)
        X_pca = pca.fit_transform(X_scaled)
        df[] = X_pca[:, ]
        df[] = X_pca[:, ]
        
         {
            : df,
            : kmeans.cluster_centers_,
            : kmeans.inertia_,
            : pca.explained_variance_ratio_
        }
    
     ():
        
        cluster_profiles = {}
        
         cluster_id  clustered_data[].unique():
            cluster_members = clustered_data[clustered_data[] == cluster_id]
            
            cluster_profiles[cluster_id] = {
                : (cluster_members),
                : cluster_members[].value_counts().to_dict(),
                : cluster_members[].mean(),
                : cluster_members[].mean(),
                : cluster_members[].mean(),
                : cluster_members.nlargest(, )[[, ]].to_dict()
            }
        
         cluster_profiles

4. Natural Language Processing (Pattern Recognition Framework)

Purpose: Extract insights from parliamentary documents, speeches, and legislative text.

CIA Platform Applications:

• Topic modeling of parliamentary motions
• Sentiment analysis of debates
• Entity extraction (policy areas, stakeholders)
• Document similarity for coalition positioning

Topic Modeling with LDA

Example: Discover Policy Topics from Parliamentary Motions

from sklearn.feature_extraction.text import CountVectorizer, TfidfVectorizer
from sklearn.decomposition import LatentDirichletAllocation
import pandas as pd

class ParliamentaryDocumentAnalyzer:
    """
    NLP analysis of parliamentary documents
    Supports: Pattern Recognition Framework
    """
    
    def __init__(self, db_connection):
        self.db = db_connection
    
    def extract_topics_from_motions(self, n_topics=10):
        """
        Latent Dirichlet Allocation (LDA) for topic discovery
        
        Data Source: document_element table (document_type = 'mot')
        Intelligence Application: Identify party policy priorities
        """
        
        # Query: Fetch parliamentary motions
        query = """
        SELECT 
            d.document_id,
            d.title,
            d.sub_title,
            d.label as motion_id,
            d.document_type,
            dpr.party,
            EXTRACT(YEAR FROM d.made_date) as year
        FROM document_element d
        JOIN document_person_reference_data dpr ON d.document_id = dpr.document_id
        WHERE d.document_type = 'mot'  -- Parliamentary motions
            AND d.made_date >= CURRENT_DATE - INTERVAL '4 years'
        """
        
        df = pd.read_sql(query, self.db)
        
        # Combine title and subtitle as document text
        df['text'] = df['title'].fillna('') + ' ' + df['sub_title'].fillna('')
        
        # Vectorize documents (Swedish stopwords)
        swedish_stopwords = ['och', 'att', 'det', 'i', 'för', , , , , , , , ]
        
        vectorizer = CountVectorizer(
            max_features=,
            stop_words=swedish_stopwords,
            ngram_range=(, ),  
            min_df=  
        )
        
        doc_term_matrix = vectorizer.fit_transform(df[])
        
        
        lda_model = LatentDirichletAllocation(
            n_components=n_topics,
            max_iter=,
            learning_method=,
            random_state=
        )
        
        lda_output = lda_model.fit_transform(doc_term_matrix)
        
        
        feature_names = vectorizer.get_feature_names_out()
        topics = {}
        
         topic_idx, topic  (lda_model.components_):
            top_words_idx = topic.argsort()[-:][::-]
            top_words = [feature_names[i]  i  top_words_idx]
            topics[] = top_words
        
        
        df[] = lda_output.argmax(axis=)
        df[] = lda_output.(axis=)
        
         {
            : topics,
            : df,
            : lda_model,
            : vectorizer
        }
    
     ():
        
        party_topic_matrix = document_topics.groupby([, ]).size().unstack(fill_value=)
        
        
        party_topic_percentage = party_topic_matrix.div(party_topic_matrix.(axis=), axis=) * 
        
         party_topic_percentage

5. Network Analysis (Network Framework)

Purpose: Model relationships between political actors as graphs to identify influence, coalitions, and power structures.

CIA Platform Applications:

• Voting alignment networks (who votes with whom)
• Co-sponsorship networks (legislative collaboration)
• Influence centrality (key power brokers)
• Coalition structure detection

NetworkX Analysis

Example: Voting Alignment Network

import networkx as nx
import pandas as pd
from scipy.stats import pearsonr

class PoliticalNetworkAnalyzer:
    """
    Graph analysis of political relationships
    Supports: Network Analysis Framework
    """
    
    def __init__(self, db_connection):
        self.db = db_connection
    
    def build_voting_alignment_network(self, threshold=0.7):
        """
        Construct weighted graph of politician voting alignment
        
        Data Source: vote_data table
        Edge weight: Pearson correlation of voting patterns
        """
        
        # Query: Pivot vote data into person x ballot matrix
        query = """
        WITH vote_matrix AS (
            SELECT 
                person_id,
                ballot_id,
                CASE vote 
                    WHEN 'Ja' THEN 1
                    WHEN 'Nej' THEN -1
                    ELSE 0
                END as vote_numeric
            FROM vote_data
            WHERE vote_date >= CURRENT_DATE - INTERVAL '12 months'
                AND vote IN ('Ja', 'Nej')
        )
        SELECT 
            vm.person_id,
            vm.ballot_id,
            vm.vote_numeric,
            p.first_name || ' ' || p.last_name as name,
            p.party
        FROM vote_matrix vm
        JOIN person_data p ON vm.person_id = p.person_id
        WHERE p.status = 'Tjänstgörande riksdagsledamot'
        """
        
        df = pd.read_sql(query, self.db)
        
        # Pivot to person x ballot matrix
        vote_pivot = df.pivot_table(
            index='person_id', 
            columns='ballot_id', 
            values='vote_numeric', 
            fill_value=0
        )
        
        # Calculate pairwise voting alignment (Pearson correlation)
        alignment_matrix = vote_pivot.T.corr()
        
        # Build NetworkX graph
        G = nx.Graph()
        
        
        person_attrs = df[[, , ]].drop_duplicates().set_index()
         person_id, attrs  person_attrs.iterrows():
            G.add_node(person_id, name=attrs[], party=attrs[])
        
        
         i, person_i  (alignment_matrix.index):
             j, person_j  (alignment_matrix.columns):
                 i < j:  
                    alignment = alignment_matrix.loc[person_i, person_j]
                     alignment >= threshold:
                        G.add_edge(person_i, person_j, weight=alignment)
        
         G
    
     ():
        
        
        
        degree_centrality = nx.degree_centrality(G)
        
        
        betweenness_centrality = nx.betweenness_centrality(G, weight=)
        
        
        eigenvector_centrality = nx.eigenvector_centrality(G, weight=, max_iter=)
        
        
        pagerank = nx.pagerank(G, weight=)
        
        
        metrics_df = pd.DataFrame({
            : (G.nodes()),
            : [G.nodes[n][]  n  G.nodes()],
            : [G.nodes[n][]  n  G.nodes()],
            : [degree_centrality[n]  n  G.nodes()],
            : [betweenness_centrality[n]  n  G.nodes()],
            : [eigenvector_centrality[n]  n  G.nodes()],
            : [pagerank[n]  n  G.nodes()]
        })
        
         metrics_df.sort_values(, ascending=)
    
     ():
        
         networkx.algorithms  community
        
        
        communities = community.greedy_modularity_communities(G, weight=)
        
        
        community_map = {}
         idx, comm  (communities):
             node  comm:
                community_map[node] = idx
        
        nx.set_node_attributes(G, community_map, )
        
         {
            : communities,
            : community.modularity(G, communities, weight=),
            : (communities)
        }

ISMS Compliance Mapping

ISO 27001:2022 Controls

A.8.16 - Monitoring Activities

• Machine learning models monitored for performance degradation
• Model retraining schedules documented and executed

A.8.32 - Change Management

• Model version control and deployment processes
• A/B testing for model updates

NIST CSF 2.0 Functions

IDENTIFY (ID)

• ID.RA-1: Asset vulnerabilities identified using predictive models
• ID.RA-3: Threats identified through anomaly detection

DETECT (DE)

• DE.AE-2: Detected events analyzed using ML classification
• DE.CM-4: Malicious code detected (adapted for behavioral anomalies)

CIS Controls v8.1

CIS Control 4: Secure Configuration of Enterprise Assets

• 4.1: Establish secure configuration process for data science pipelines
  • Model hyperparameters documented and version-controlled

CIS Control 12: Network Infrastructure Management

• 12.8: Establish and maintain dedicated infrastructure for machine learning workloads

Hack23 ISMS Policy References

Data Classification Policy

• Link: https://github.com/Hack23/ISMS-PUBLIC/blob/main/Data_Classification_Policy.md
• Application: Training data classified, model outputs marked with confidence levels

AI Policy

• Link: https://github.com/Hack23/ISMS-PUBLIC/blob/main/AI_Policy.md
• Application: ML models follow AI governance framework, bias mitigation procedures

Secure Development Policy

• Link: https://github.com/Hack23/ISMS-PUBLIC/blob/main/Secure_Development_Policy.md
• Application: Model development follows SDLC, security testing integrated

References

Official Documentation:

• Scikit-learn: https://scikit-learn.org/
• Statsmodels: https://www.statsmodels.org/
• NetworkX: https://networkx.org/
• NLTK: https://www.nltk.org/

CIA Platform Documentation:

• Data Analysis: DATA_ANALYSIS_INTOP_OSINT.md
• Database Views: DATABASE_VIEW_INTELLIGENCE_CATALOG.md
• Risk Rules: RISK_RULES_INTOP_OSINT.md

Academic Sources:

• "Pattern Recognition and Machine Learning" - Christopher Bishop
• "Introduction to Statistical Learning" - James, Witten, Hastie, Tibshirani
• "Network Science" - Albert-László Barabási

---

🔗 Integration with agentic workflows & analysis artifacts

This skill is consumed by the 11 agentic news workflows in .github/workflows/news-*.md. The authoritative contract lives in .github/prompts/README.md; this skill supplies domain expertise on top of that contract.

• Analysis product → ai-driven-analysis-guide.md + every template in analysis/templates/.
• Required before any article: 9 core artifacts (14 for Tier-C) in analysis/daily/$ARTICLE_DATE/$SUBFOLDER/; 05-analysis-gate.md is the single blocking gate.
• gh-aw v0.69.3 docs: abridged · complete · blog series.

---

🌐 IMF as Primary Source for Cross-Country Economic Panels

Effective: 2026-04-24

IMF data quality controls in statistical analysis

Control	Implementation
Vintage discipline	Tag every IMF observation with vintage_label (e.g., 2026-04 for April WEO); never silently overwrite
Supersedes-chain	When new vintage publishes, link via supersedes pointer; preserves audit trail
Methodology consistency	SNA 2008 / GFSM 2014 / BPM6 — uniform across all member countries
Projection-vs-observation flag	is_projection: true/false on every WEO/FM observation
Cross-validation against SCB	For Sweden-specific quarterly figures, IMF + SCB compared; >0.3pp delta opens issue

Statistical pipelines using IMF data

• Cross-country panel regressions — IMF WEO provides balanced panel (190+ countries × 50+ years) on uniform methodology
• Time-series forecasting — IMF WEO + FM projections (T+5) feed scenario analysis; vintage-stamped for backtesting
• Bilateral trade network analysis — IMF DOTS provides directed bilateral flows (monthly) for Swedish trade-partner network
• Commodity-price pass-through models — IMF PCPS provides canonical price benchmarks
• Fiscal-multiplier estimation — IMF GFS_COFOG provides function-level spending decomposition

Canonical data-science rule: Every statistical analysis on economic data uses IMF as the primary source; vintage labels preserved through analytical pipeline; provenance recorded in output. See analysis/imf/data-dictionary.md and .github/aw/ECONOMIC_DATA_CONTRACT.md v2.1.