Upload 4 files

app.py

@@ -1,758 +1,300 @@
-import os
-import re
-import json
-import time
-import traceback
+import os, re, json, time, traceback
 from pathlib import Path
 from typing import Dict, Any, List, Tuple
-
 import pandas as pd
 import gradio as gr
 import papermill as pm
 import plotly.graph_objects as go
 
-# Optional LLM (HuggingFace Inference API)
 try:
     from huggingface_hub import InferenceClient
 except Exception:
     InferenceClient = None
 
-# =========================================================
-# CONFIG
-# =========================================================
-
 BASE_DIR = Path(__file__).resolve().parent
-
 NB1 = os.environ.get("NB1", "datacreation.ipynb").strip()
 NB2 = os.environ.get("NB2", "pythonanalysis.ipynb").strip()
-
-RUNS_DIR = BASE_DIR / "runs"
-ART_DIR = BASE_DIR / "artifacts"
+RUNS_DIR   = BASE_DIR / "runs"
+ART_DIR    = BASE_DIR / "artifacts"
 PY_FIG_DIR = ART_DIR / "py" / "figures"
 PY_TAB_DIR = ART_DIR / "py" / "tables"
-
 PAPERMILL_TIMEOUT = int(os.environ.get("PAPERMILL_TIMEOUT", "1800"))
-MAX_PREVIEW_ROWS = int(os.environ.get("MAX_FILE_PREVIEW_ROWS", "50"))
-MAX_LOG_CHARS = int(os.environ.get("MAX_LOG_CHARS", "8000"))
-
-HF_API_KEY = os.environ.get("HF_API_KEY", "").strip()
-MODEL_NAME = os.environ.get("MODEL_NAME", "deepseek-ai/DeepSeek-R1").strip()
-HF_PROVIDER = os.environ.get("HF_PROVIDER", "novita").strip()
-N8N_WEBHOOK_URL = os.environ.get("N8N_WEBHOOK_URL", "").strip()
-
+MAX_PREVIEW_ROWS  = int(os.environ.get("MAX_FILE_PREVIEW_ROWS", "50"))
+HF_API_KEY        = os.environ.get("HF_API_KEY", "").strip()
+MODEL_NAME        = os.environ.get("MODEL_NAME", "deepseek-ai/DeepSeek-R1").strip()
+HF_PROVIDER       = os.environ.get("HF_PROVIDER", "novita").strip()
+N8N_WEBHOOK_URL   = os.environ.get("N8N_WEBHOOK_URL", "").strip()
+ANTHROPIC_API_KEY = os.environ.get("ANTHROPIC_API_KEY", "").strip()
 LLM_ENABLED = bool(HF_API_KEY) and InferenceClient is not None
-llm_client = (
-    InferenceClient(provider=HF_PROVIDER, api_key=HF_API_KEY)
-    if LLM_ENABLED
-    else None
-)
-
-# =========================================================
-# HELPERS
-# =========================================================
+llm_client  = InferenceClient(provider=HF_PROVIDER, api_key=HF_API_KEY) if LLM_ENABLED else None
 
 def ensure_dirs():
     for p in [RUNS_DIR, ART_DIR, PY_FIG_DIR, PY_TAB_DIR]:
         p.mkdir(parents=True, exist_ok=True)
 
-def stamp():
-    return time.strftime("%Y%m%d-%H%M%S")
+def _ls(d, exts):
+    return sorted(p.name for p in d.iterdir() if p.is_file() and p.suffix.lower() in exts) if d.is_dir() else []
 
-def tail(text: str, n: int = MAX_LOG_CHARS) -> str:
-    return (text or "")[-n:]
+def _read_csv(p): return pd.read_csv(p, nrows=MAX_PREVIEW_ROWS)
+def _read_json(p):
+    with open(p, encoding='utf-8') as f: return json.load(f)
 
-def _ls(dir_path: Path, exts: Tuple[str, ...]) -> List[str]:
-    if not dir_path.is_dir():
-        return []
-    return sorted(p.name for p in dir_path.iterdir() if p.is_file() and p.suffix.lower() in exts)
+def artifacts_index():
+    return {"python": {"figures": _ls(PY_FIG_DIR, (".png",".jpg")), "tables": _ls(PY_TAB_DIR, (".csv",".json"))}}
 
-def _read_csv(path: Path) -> pd.DataFrame:
-    return pd.read_csv(path, nrows=MAX_PREVIEW_ROWS)
-
-def _read_json(path: Path):
-    with path.open(encoding="utf-8") as f:
-        return json.load(f)
-
-def artifacts_index() -> Dict[str, Any]:
-    return {
-        "python": {
-            "figures": _ls(PY_FIG_DIR, (".png", ".jpg", ".jpeg")),
-            "tables": _ls(PY_TAB_DIR, (".csv", ".json")),
-        },
-    }
+def load_kpis():
+    c = PY_TAB_DIR / "kpis.json"
+    if c.exists():
+        try: return _read_json(c)
+        except: pass
+    return {}
 
-# =========================================================
-# PIPELINE RUNNERS
-# =========================================================
+def _load_table_safe(p):
+    try:
+        if p.suffix == ".json":
+            obj = _read_json(p)
+            return pd.DataFrame([obj] if isinstance(obj, dict) else obj)
+        return _read_csv(p)
+    except Exception as e:
+        return pd.DataFrame([{"error": str(e)}])
 
-def run_notebook(nb_name: str) -> str:
+def run_notebook(nb):
     ensure_dirs()
-    nb_in = BASE_DIR / nb_name
-    if not nb_in.exists():
-        return f"ERROR: {nb_name} not found."
-    nb_out = RUNS_DIR / f"run_{stamp()}_{nb_name}"
-    pm.execute_notebook(
-        input_path=str(nb_in),
-        output_path=str(nb_out),
-        cwd=str(BASE_DIR),
-        log_output=True,
-        progress_bar=False,
-        request_save_on_cell_execute=True,
-        execution_timeout=PAPERMILL_TIMEOUT,
-    )
-    return f"Executed {nb_name}"
-
-
-def run_datacreation() -> str:
+    nb_in = BASE_DIR / nb
+    if not nb_in.exists(): return f"ERROR: {nb} not found."
+    nb_out = RUNS_DIR / f"run_{time.strftime('%Y%m%d-%H%M%S')}_{nb}"
+    pm.execute_notebook(str(nb_in), str(nb_out), cwd=str(BASE_DIR),
+                        log_output=True, progress_bar=False, execution_timeout=PAPERMILL_TIMEOUT)
+    return f"Executed {nb}"
+
+def run_datacreation():
     try:
         log = run_notebook(NB1)
         csvs = [f.name for f in BASE_DIR.glob("*.csv")]
-        return f"OK  {log}\n\nCSVs now in /app:\n" + "\n".join(f"  - {c}" for c in sorted(csvs))
+        return f"OK  {log}\n\nCSVs:\n" + "\n".join(f"  - {c}" for c in sorted(csvs))
     except Exception as e:
         return f"FAILED  {e}\n\n{traceback.format_exc()[-2000:]}"
 
-
-def run_pythonanalysis() -> str:
+def run_pythonanalysis():
     try:
         log = run_notebook(NB2)
         idx = artifacts_index()
-        figs = idx["python"]["figures"]
-        tabs = idx["python"]["tables"]
-        return (
-            f"OK  {log}\n\n"
-            f"Figures: {', '.join(figs) or '(none)'}\n"
-            f"Tables:  {', '.join(tabs) or '(none)'}"
-        )
+        return (f"OK  {log}\n\nFigures: {', '.join(idx['python']['figures']) or '(none)'}\n"
+                f"Tables: {', '.join(idx['python']['tables']) or '(none)'}")
     except Exception as e:
         return f"FAILED  {e}\n\n{traceback.format_exc()[-2000:]}"
 
-
-def run_full_pipeline() -> str:
-    logs = []
-    logs.append("=" * 50)
-    logs.append("STEP 1/2: Data Creation (web scraping + synthetic data)")
-    logs.append("=" * 50)
-    logs.append(run_datacreation())
-    logs.append("")
-    logs.append("=" * 50)
-    logs.append("STEP 2/2: Python Analysis (sentiment, ARIMA, dashboard)")
-    logs.append("=" * 50)
-    logs.append(run_pythonanalysis())
-    return "\n".join(logs)
-
-
-# =========================================================
-# GALLERY LOADERS
-# =========================================================
-
-def _load_all_figures() -> List[Tuple[str, str]]:
-    """Return list of (filepath, caption) for Gallery."""
-    items = []
-    for p in sorted(PY_FIG_DIR.glob("*.png")):
-        items.append((str(p), p.stem.replace('_', ' ').title()))
-    return items
-
-
-def _load_table_safe(path: Path) -> pd.DataFrame:
+def run_full_pipeline():
+    return "\n".join(["="*50,"STEP 1/2: Data Collection","="*50,run_datacreation(),"",
+                        "="*50,"STEP 2/2: Analysis","="*50,run_pythonanalysis()])
+
+def _call_anthropic(system, messages, max_tokens=700):
+    if not ANTHROPIC_API_KEY: return None
+    import urllib.request
+    payload = json.dumps({"model":"claude-sonnet-4-20250514","max_tokens":max_tokens,
+                           "system":system,"messages":messages}).encode()
+    req = urllib.request.Request("https://api.anthropic.com/v1/messages", data=payload,
+        headers={"Content-Type":"application/json","x-api-key":ANTHROPIC_API_KEY,
+                 "anthropic-version":"2023-06-01"}, method="POST")
     try:
-        if path.suffix == ".json":
-            obj = _read_json(path)
-            if isinstance(obj, dict):
-                return pd.DataFrame([obj])
-            return pd.DataFrame(obj)
-        return _read_csv(path)
-    except Exception as e:
-        return pd.DataFrame([{"error": str(e)}])
-
-
-def refresh_gallery():
-    """Called when user clicks Refresh on Gallery tab."""
-    figures = _load_all_figures()
-    idx = artifacts_index()
-
-    table_choices = list(idx["python"]["tables"])
-
-    default_df = pd.DataFrame()
-    if table_choices:
-        default_df = _load_table_safe(PY_TAB_DIR / table_choices[0])
-
-    return (
-        figures if figures else [],
-        gr.update(choices=table_choices, value=table_choices[0] if table_choices else None),
-        default_df,
-    )
-
-
-def on_table_select(choice: str):
-    if not choice:
-        return pd.DataFrame([{"hint": "Select a table above."}])
-    path = PY_TAB_DIR / choice
-    if not path.exists():
-        return pd.DataFrame([{"error": f"File not found: {choice}"}])
-    return _load_table_safe(path)
-
-
-# =========================================================
-# KPI LOADER
-# =========================================================
-
-def load_kpis() -> Dict[str, Any]:
-    for candidate in [PY_TAB_DIR / "kpis.json", PY_FIG_DIR / "kpis.json"]:
-        if candidate.exists():
-            try:
-                return _read_json(candidate)
-            except Exception:
-                pass
-    return {}
-
-
-# =========================================================
-# AI DASHBOARD -- LLM picks what to display
-# =========================================================
-
-DASHBOARD_SYSTEM = """You are an AI dashboard assistant for a book-sales analytics app.
-The user asks questions or requests about their data. You have access to pre-computed
-artifacts from a Python analysis pipeline.
-
-AVAILABLE ARTIFACTS (only reference ones that exist):
-{artifacts_json}
-
+        with urllib.request.urlopen(req, timeout=30) as r:
+            return json.loads(r.read())["content"][0]["text"]
+    except Exception as e: return f"Anthropic error: {e}"
+
+SYSTEM = """You are an AI portfolio analyst for an emerging market investment fund.
+Help fund managers decide which countries to overweight or underweight during geopolitical stress.
+You have access to World Bank macro data, synthetic geopolitical risk scores, VADER news sentiment, and Random Forest investment signals.
+AVAILABLE ARTIFACTS: {artifacts_json}
 KPI SUMMARY: {kpis_json}
+Answer in 2-4 sentences. End with: ```json {{"show": "figure"|"table"|"none", "scope":"python","filename":"..."}} ```
+ROUTING: GDP/trend->df_dashboard.csv or gdp_heatmap.png | risk->geo_risk_heatmap.png | sentiment->vader_by_country.csv | signal->investment_signal_summary.csv | arima->arima_gdp_forecast.csv | rf->rf_feature_importance.png | predictions->country_predictions_latest.csv | overview->kpis.json"""
 
-YOUR JOB:
-1. Answer the user's question conversationally using the KPIs and your knowledge of the artifacts.
-2. At the END of your response, output a JSON block (fenced with ```json ... ```) that tells
-   the dashboard which artifact to display. The JSON must have this shape:
-   {{"show": "figure"|"table"|"none", "scope": "python", "filename": "..."}}
-
-   - Use "show": "figure" to display a chart image.
-   - Use "show": "table" to display a CSV/JSON table.
-   - Use "show": "none" if no artifact is relevant.
-
-RULES:
-- If the user asks about sales trends or forecasting by title, show sales_trends or arima figures.
-- If the user asks about sentiment, show sentiment figure or sentiment_counts table.
-- If the user asks about forecast accuracy or ARIMA, show arima figures.
-- If the user asks about top sellers, show top_titles_by_units_sold.csv.
-- If the user asks a general data question, pick the most relevant artifact.
-- Keep your answer concise (2-4 sentences), then the JSON block.
-"""
+JSON_RE = re.compile(r"```json\s*(\{.*?\})\s*```", re.DOTALL)
+FB_RE   = re.compile(r"\{[^{}]*\"show\"[^{}]*\}", re.DOTALL)
 
-JSON_BLOCK_RE = re.compile(r"```json\s*(\{.*?\})\s*```", re.DOTALL)
-FALLBACK_JSON_RE = re.compile(r"\{[^{}]*\"show\"[^{}]*\}", re.DOTALL)
-
-
-def _parse_display_directive(text: str) -> Dict[str, str]:
-    m = JSON_BLOCK_RE.search(text)
+def _parse_directive(text):
+    m = JSON_RE.search(text)
     if m:
-        try:
-            return json.loads(m.group(1))
-        except json.JSONDecodeError:
-            pass
-    m = FALLBACK_JSON_RE.search(text)
+        try: return json.loads(m.group(1))
+        except: pass
+    m = FB_RE.search(text)
     if m:
-        try:
-            return json.loads(m.group(0))
-        except json.JSONDecodeError:
-            pass
-    return {"show": "none"}
+        try: return json.loads(m.group(0))
+        except: pass
+    return {"show":"none"}
 
+def _clean(text): return JSON_RE.sub("", text).strip()
 
-def _clean_response(text: str) -> str:
-    """Strip the JSON directive block from the displayed response."""
-    return JSON_BLOCK_RE.sub("", text).strip()
-
-
-def _n8n_call(msg: str) -> Tuple[str, Dict]:
-    """Call the student's n8n webhook and return (reply, directive)."""
-    import requests as req
-    try:
-        resp = req.post(N8N_WEBHOOK_URL, json={"question": msg}, timeout=20)
-        data = resp.json()
-        answer = data.get("answer", "No response from n8n workflow.")
-        chart = data.get("chart", "none")
-        if chart and chart != "none":
-            return answer, {"show": "figure", "chart": chart}
-        return answer, {"show": "none"}
-    except Exception as e:
-        return f"n8n error: {e}. Falling back to keyword matching.", None
-
-
-def ai_chat(user_msg: str, history: list):
-    """Chat function for the AI Dashboard tab."""
-    if not user_msg or not user_msg.strip():
-        return history, "", None, None
-
-    idx = artifacts_index()
-    kpis = load_kpis()
-
-    # Priority: n8n webhook > HF LLM > keyword fallback
+def _keyword(msg, idx, kpis):
+    ml = msg.lower()
+    kpi_txt = f"Summary: {kpis.get('Countries Analysed','?')} countries, avg GDP {kpis.get('Avg GDP Growth','?')}, RF accuracy {kpis.get('RF Accuracy','?')}." if kpis else ""
+    if not idx["python"]["figures"] and not idx["python"]["tables"]:
+        return "No data yet — run the pipeline first.", {"show":"none"}
+    if any(w in ml for w in ["gdp","growth","macro","trend"]): return f"Here is the GDP growth trend. {kpi_txt}", {"show":"figure","filename":"gdp_heatmap.png"}
+    if any(w in ml for w in ["risk","geopolit","conflict","stress"]): return f"Here is the geo risk heatmap. {kpi_txt}", {"show":"figure","filename":"geo_risk_heatmap.png"}
+    if any(w in ml for w in ["sentiment","vader","news","headline"]): return f"Here is the VADER sentiment. {kpi_txt}", {"show":"figure","filename":"vader_sentiment.png"}
+    if any(w in ml for w in ["overweight","underweight","signal","invest","portfolio","allocat"]): return f"Here are investment signals. {kpi_txt}", {"show":"figure","filename":"investment_signal.png"}
+    if any(w in ml for w in ["arima","forecast","predict","future"]): return f"Here is the ARIMA forecast. {kpi_txt}", {"show":"figure","filename":"arima_gdp_forecast.png"}
+    if any(w in ml for w in ["random forest","rf","feature","importance","classif"]): return f"Here are RF importances. {kpi_txt}", {"show":"figure","filename":"rf_feature_importance.png"}
+    if any(w in ml for w in ["country","which","rank","top","best","worst"]): return f"Here are country predictions. {kpi_txt}", {"show":"table","scope":"python","filename":"country_predictions_latest.csv"}
+    if any(w in ml for w in ["dashboard","overview","summary","kpi"]): return f"Dashboard. {kpi_txt}", {"show":"table","scope":"python","filename":"kpis.json"}
+    return f"Ask about: GDP trends, geo risk, sentiment, investment signals, ARIMA forecasts, RF features, or country predictions. {kpi_txt}", {"show":"none"}
+
+def ai_chat(user_msg, history):
+    if not user_msg or not user_msg.strip(): return history, "", None, None
+    idx  = artifacts_index(); kpis = load_kpis()
     if N8N_WEBHOOK_URL:
-        reply, directive = _n8n_call(user_msg)
-        if directive is None:
-            reply_fb, directive = _keyword_fallback(user_msg, idx, kpis)
-            reply += "\n\n" + reply_fb
-    elif not LLM_ENABLED:
-        reply, directive = _keyword_fallback(user_msg, idx, kpis)
-    else:
-        system = DASHBOARD_SYSTEM.format(
-            artifacts_json=json.dumps(idx, indent=2),
-            kpis_json=json.dumps(kpis, indent=2) if kpis else "(no KPIs yet, run the pipeline first)",
-        )
-        msgs = [{"role": "system", "content": system}]
-        for entry in (history or [])[-6:]:
-            msgs.append(entry)
-        msgs.append({"role": "user", "content": user_msg})
-
+        import requests as req
         try:
-            r = llm_client.chat_completion(
-                model=MODEL_NAME,
-                messages=msgs,
-                temperature=0.3,
-                max_tokens=600,
-                stream=False,
-            )
-            raw = (
-                r["choices"][0]["message"]["content"]
-                if isinstance(r, dict)
-                else r.choices[0].message.content
-            )
-            directive = _parse_display_directive(raw)
-            reply = _clean_response(raw)
+            r = req.post(N8N_WEBHOOK_URL, json={"question":user_msg}, timeout=20); d = r.json()
+            reply = d.get("answer","No n8n response."); chart = d.get("chart","none")
+            directive = {"show":"figure","chart":chart} if chart and chart!="none" else {"show":"none"}
         except Exception as e:
-            reply = f"LLM error: {e}. Falling back to keyword matching."
-            reply_fb, directive = _keyword_fallback(user_msg, idx, kpis)
-            reply += "\n\n" + reply_fb
-
-    # Resolve artifacts — build interactive Plotly charts when possible
-    chart_out = None
-    tab_out = None
-    show = directive.get("show", "none")
-    fname = directive.get("filename", "")
-    chart_name = directive.get("chart", "")
-
-    # Interactive chart builders keyed by name
-    chart_builders = {
-        "sales": build_sales_chart,
-        "sentiment": build_sentiment_chart,
-        "top_sellers": build_top_sellers_chart,
-    }
-
-    if chart_name and chart_name in chart_builders:
-        chart_out = chart_builders[chart_name]()
-    elif show == "figure" and fname:
-        # Fallback: try to match filename to a chart builder
-        if "sales_trend" in fname:
-            chart_out = build_sales_chart()
-        elif "sentiment" in fname:
-            chart_out = build_sentiment_chart()
-        elif "arima" in fname or "forecast" in fname:
-            chart_out = build_sales_chart()  # closest interactive equivalent
-        else:
-            chart_out = _empty_chart(f"No interactive chart for {fname}")
-
-    if show == "table" and fname:
-        fp = PY_TAB_DIR / fname
-        if fp.exists():
-            tab_out = _load_table_safe(fp)
+            reply = f"n8n error: {e}"; rb, directive = _keyword(user_msg, idx, kpis); reply += "\n\n"+rb
+    elif ANTHROPIC_API_KEY:
+        system = SYSTEM.format(artifacts_json=json.dumps(idx,indent=2), kpis_json=json.dumps(kpis,indent=2) if kpis else "(run pipeline first)")
+        msgs = list(history or [])[-6:] + [{"role":"user","content":user_msg}]
+        raw = _call_anthropic(system, msgs)
+        if raw and "error" not in (raw or "")[:20]:
+            directive = _parse_directive(raw); reply = _clean(raw)
         else:
-            reply += f"\n\n*(Could not find table: {fname})*"
-
-    new_history = (history or []) + [
-        {"role": "user", "content": user_msg},
-        {"role": "assistant", "content": reply},
-    ]
-
-    return new_history, "", chart_out, tab_out
-
-
-def _keyword_fallback(msg: str, idx: Dict, kpis: Dict) -> Tuple[str, Dict]:
-    """Simple keyword matcher when LLM is unavailable."""
-    msg_lower = msg.lower()
-
-    if not idx["python"]["figures"] and not idx["python"]["tables"]:
-        return (
-            "No artifacts found yet. Please run the pipeline first (Tab 1), "
-            "then come back here to explore the results.",
-            {"show": "none"},
-        )
-
-    kpi_text = ""
-    if kpis:
-        total = kpis.get("total_units_sold", 0)
-        kpi_text = (
-            f"Quick summary: **{kpis.get('n_titles', '?')}** book titles across "
-            f"**{kpis.get('n_months', '?')}** months, with **{total:,.0f}** total units sold."
-        )
-
-    if any(w in msg_lower for w in ["trend", "sales trend", "monthly sale"]):
-        return (
-            f"Here are the sales trends. {kpi_text}",
-            {"show": "figure", "chart": "sales"},
-        )
-
-    if any(w in msg_lower for w in ["sentiment", "review", "positive", "negative"]):
-        return (
-            f"Here is the sentiment distribution across sampled book titles. {kpi_text}",
-            {"show": "figure", "chart": "sentiment"},
-        )
-
-    if any(w in msg_lower for w in ["arima", "forecast", "predict"]):
-        return (
-            f"Here are the sales trends and forecasts. {kpi_text}",
-            {"show": "figure", "chart": "sales"},
-        )
-
-    if any(w in msg_lower for w in ["top", "best sell", "popular", "rank"]):
-        return (
-            f"Here are the top-selling titles by units sold. {kpi_text}",
-            {"show": "table", "scope": "python", "filename": "top_titles_by_units_sold.csv"},
-        )
-
-    if any(w in msg_lower for w in ["price", "pricing", "decision"]):
-        return (
-            f"Here are the pricing decisions. {kpi_text}",
-            {"show": "table", "scope": "python", "filename": "pricing_decisions.csv"},
-        )
-
-    if any(w in msg_lower for w in ["dashboard", "overview", "summary", "kpi"]):
-        return (
-            f"Dashboard overview: {kpi_text}\n\nAsk me about sales trends, sentiment, forecasts, "
-            "pricing, or top sellers to see specific visualizations.",
-            {"show": "table", "scope": "python", "filename": "df_dashboard.csv"},
-        )
-
-    # Default
-    return (
-        f"I can show you various analyses. {kpi_text}\n\n"
-        "Try asking about: **sales trends**, **sentiment**, **ARIMA forecasts**, "
-        "**pricing decisions**, **top sellers**, or **dashboard overview**.",
-        {"show": "none"},
-    )
-
-
-# =========================================================
-# KPI CARDS (BubbleBusters style)
-# =========================================================
-
-def render_kpi_cards() -> str:
-    kpis = load_kpis()
-    if not kpis:
-        return (
-            '<div style="background:rgba(255,255,255,.65);backdrop-filter:blur(16px);'
-            'border-radius:20px;padding:28px;text-align:center;'
-            'border:1.5px solid rgba(255,255,255,.7);'
-            'box-shadow:0 8px 32px rgba(124,92,191,.08);">'
-            '<div style="font-size:36px;margin-bottom:10px;">📊</div>'
-            '<div style="color:#a48de8;font-size:14px;'
-            'font-weight:800;margin-bottom:6px;">No data yet</div>'
-            '<div style="color:#9d8fc4;font-size:12px;">'
-            'Run the pipeline to populate these cards.</div>'
-            '</div>'
-        )
-
-    def card(icon, label, value, colour):
-        return f"""
-        <div style="background:rgba(255,255,255,.72);backdrop-filter:blur(16px);
-                    border-radius:20px;padding:18px 14px 16px;text-align:center;
-                    border:1.5px solid rgba(255,255,255,.8);
-                    box-shadow:0 4px 16px rgba(124,92,191,.08);
-                    border-top:3px solid {colour};">
-            <div style="font-size:26px;margin-bottom:7px;line-height:1;">{icon}</div>
-            <div style="color:#9d8fc4;font-size:9.5px;text-transform:uppercase;
-                        letter-spacing:1.8px;margin-bottom:7px;font-weight:800;">{label}</div>
-            <div style="color:#2d1f4e;font-size:16px;font-weight:800;">{value}</div>
-        </div>"""
-
-    kpi_config = [
-        ("n_titles",         "📚", "Book Titles",  "#a48de8"),
-        ("n_months",         "📅", "Time Periods", "#7aa6f8"),
-        ("total_units_sold", "📦", "Units Sold",   "#6ee7c7"),
-        ("total_revenue",    "💰", "Revenue",      "#3dcba8"),
-    ]
-
-    html = (
-        '<div style="display:grid;grid-template-columns:repeat(auto-fit,minmax(140px,1fr));'
-        'gap:12px;margin-bottom:24px;">'
-    )
-    for key, icon, label, colour in kpi_config:
-        val = kpis.get(key)
-        if val is None:
-            continue
-        if isinstance(val, (int, float)) and val > 100:
-            val = f"{val:,.0f}"
-        html += card(icon, label, str(val), colour)
-    # Extra KPIs not in config
-    known = {k for k, *_ in kpi_config}
-    for key, val in kpis.items():
-        if key not in known:
-            label = key.replace("_", " ").title()
-            if isinstance(val, (int, float)) and val > 100:
-                val = f"{val:,.0f}"
-            html += card("📈", label, str(val), "#8fa8f8")
-    html += "</div>"
-    return html
-
-
-# =========================================================
-# INTERACTIVE PLOTLY CHARTS (BubbleBusters style)
-# =========================================================
-
-CHART_PALETTE = ["#7c5cbf", "#2ec4a0", "#e8537a", "#e8a230", "#5e8fef",
-                 "#c45ea8", "#3dbacc", "#a0522d", "#6aaa3a", "#d46060"]
-
-def _styled_layout(**kwargs) -> dict:
-    defaults = dict(
-        template="plotly_white",
-        paper_bgcolor="rgba(255,255,255,0.95)",
-        plot_bgcolor="rgba(255,255,255,0.98)",
-        font=dict(family="system-ui, sans-serif", color="#2d1f4e", size=12),
-        margin=dict(l=60, r=20, t=70, b=70),
-        legend=dict(
-            orientation="h", yanchor="bottom", y=1.02, xanchor="right", x=1,
-            bgcolor="rgba(255,255,255,0.92)",
-            bordercolor="rgba(124,92,191,0.35)", borderwidth=1,
-        ),
-        title=dict(font=dict(size=15, color="#4b2d8a")),
-    )
-    defaults.update(kwargs)
-    return defaults
-
-
-def _empty_chart(title: str) -> go.Figure:
-    fig = go.Figure()
-    fig.update_layout(
-        title=title, height=420, template="plotly_white",
-        paper_bgcolor="rgba(255,255,255,0.95)",
-        annotations=[dict(text="Run the pipeline to generate data",
-            x=0.5, y=0.5, xref="paper", yref="paper", showarrow=False,
-            font=dict(size=14, color="rgba(124,92,191,0.5)"))],
-    )
-    return fig
-
-
-def build_sales_chart() -> go.Figure:
-    path = PY_TAB_DIR / "df_dashboard.csv"
-    if not path.exists():
-        return _empty_chart("Sales Trends — run the pipeline first")
-    df = pd.read_csv(path)
-    date_col = next((c for c in df.columns if "month" in c.lower() or "date" in c.lower()), None)
-    val_cols = [c for c in df.columns if c != date_col and df[c].dtype in ("float64", "int64")]
-    if not date_col or not val_cols:
-        return _empty_chart("Could not auto-detect columns in df_dashboard.csv")
-    df[date_col] = pd.to_datetime(df[date_col], errors="coerce")
-    fig = go.Figure()
-    for i, col in enumerate(val_cols):
-        fig.add_trace(go.Scatter(
-            x=df[date_col], y=df[col], name=col.replace("_", " ").title(),
-            mode="lines+markers", line=dict(color=CHART_PALETTE[i % len(CHART_PALETTE)], width=2),
-            marker=dict(size=4),
-            hovertemplate=f"<b>{col.replace('_',' ').title()}</b><br>%{{x|%b %Y}}: %{{y:,.0f}}<extra></extra>",
-        ))
-    fig.update_layout(**_styled_layout(height=450, hovermode="x unified",
-                                        title=dict(text="Monthly Overview")))
-    fig.update_xaxes(gridcolor="rgba(124,92,191,0.15)", showgrid=True)
-    fig.update_yaxes(gridcolor="rgba(124,92,191,0.15)", showgrid=True)
-    return fig
-
-
-def build_sentiment_chart() -> go.Figure:
-    path = PY_TAB_DIR / "sentiment_counts_sampled.csv"
-    if not path.exists():
-        return _empty_chart("Sentiment Distribution — run the pipeline first")
-    df = pd.read_csv(path)
-    title_col = df.columns[0]
-    sent_cols = [c for c in ["negative", "neutral", "positive"] if c in df.columns]
-    if not sent_cols:
-        return _empty_chart("No sentiment columns found in CSV")
-    colors = {"negative": "#e8537a", "neutral": "#5e8fef", "positive": "#2ec4a0"}
-    fig = go.Figure()
-    for col in sent_cols:
-        fig.add_trace(go.Bar(
-            name=col.title(), y=df[title_col], x=df[col],
-            orientation="h", marker_color=colors.get(col, "#888"),
-            hovertemplate=f"<b>{col.title()}</b>: %{{x}}<extra></extra>",
-        ))
-    fig.update_layout(**_styled_layout(
-        height=max(400, len(df) * 28), barmode="stack",
-        title=dict(text="Sentiment Distribution by Book"),
-    ))
-    fig.update_xaxes(title="Number of Reviews")
-    fig.update_yaxes(autorange="reversed")
-    return fig
-
-
-def build_top_sellers_chart() -> go.Figure:
-    path = PY_TAB_DIR / "top_titles_by_units_sold.csv"
-    if not path.exists():
-        return _empty_chart("Top Sellers — run the pipeline first")
-    df = pd.read_csv(path).head(15)
-    title_col = next((c for c in df.columns if "title" in c.lower()), df.columns[0])
-    val_col = next((c for c in df.columns if "unit" in c.lower() or "sold" in c.lower()), df.columns[-1])
-    fig = go.Figure(go.Bar(
-        y=df[title_col], x=df[val_col], orientation="h",
-        marker=dict(color=df[val_col], colorscale=[[0, "#c5b4f0"], [1, "#7c5cbf"]]),
-        hovertemplate="<b>%{y}</b><br>Units: %{x:,.0f}<extra></extra>",
-    ))
-    fig.update_layout(**_styled_layout(
-        height=max(400, len(df) * 30),
-        title=dict(text="Top Selling Titles"), showlegend=False,
-    ))
-    fig.update_yaxes(autorange="reversed")
-    fig.update_xaxes(title="Total Units Sold")
-    return fig
-
-
-def refresh_dashboard():
-    return render_kpi_cards(), build_sales_chart(), build_sentiment_chart(), build_top_sellers_chart()
-
-
-# =========================================================
-# UI
-# =========================================================
+            reply = raw or "API unavailable."; rb, directive = _keyword(user_msg, idx, kpis); reply += "\n\n"+rb
+    elif LLM_ENABLED:
+        system = SYSTEM.format(artifacts_json=json.dumps(idx,indent=2), kpis_json=json.dumps(kpis,indent=2) if kpis else "(run pipeline first)")
+        try:
+            r = llm_client.chat_completion(model=MODEL_NAME,
+                messages=[{"role":"system","content":system}]+(history or [])[-6:]+[{"role":"user","content":user_msg}],
+                temperature=0.3,max_tokens=600,stream=False)
+            raw = r["choices"][0]["message"]["content"] if isinstance(r,dict) else r.choices[0].message.content
+            directive = _parse_directive(raw); reply = _clean(raw)
+        except Exception as e:
+            reply = f"LLM error: {e}."; rb, directive = _keyword(user_msg, idx, kpis); reply += "\n\n"+rb
+    else:
+        reply, directive = _keyword(user_msg, idx, kpis)
+
+    show=directive.get("show","none"); fname=directive.get("filename","")
+    chart_out=tab_out=None
+    CHART_MAP={"investment_signal.png":build_signal_chart,"vader_sentiment.png":build_sentiment_chart,"vader_by_country.csv":build_sentiment_chart,"gdp_heatmap.png":build_gdp_trend_chart}
+    if show=="figure":
+        builder=CHART_MAP.get(fname)
+        if builder: chart_out=builder()
+        elif "arima" in fname or "forecast" in fname: chart_out=build_arima_chart()
+        else: chart_out=_empty_chart(fname or "Chart")
+    elif show=="table" and fname:
+        fp=PY_TAB_DIR/fname
+        if fp.exists(): tab_out=_load_table_safe(fp)
+        else: reply+=f"\n\n*(Table not found: {fname})*"
+    new_hist=(history or [])+[{"role":"user","content":user_msg},{"role":"assistant","content":reply}]
+    return new_hist,"",chart_out,tab_out
+
+def render_kpi_cards():
+    kpis=load_kpis()
+    if not kpis: return '<div style="text-align:center;padding:28px;background:rgba(255,255,255,.65);border-radius:20px"><div style="font-size:36px">🌍</div><div style="color:#a48de8;font-weight:800">No data yet</div><div style="color:#9d8fc4;font-size:12px">Run the pipeline first.</div></div>'
+    icons={"Countries Analysed":"🌍","Years Covered":"📅","Avg GDP Growth":"📈","Avg Geo Risk":"⚠️","Top Overweight":"✅","Top Underweight":"🔻","RF Accuracy":"🤖","Headlines Analysed":"📰"}
+    colours={"Countries Analysed":"#a48de8","Years Covered":"#7aa6f8","Avg GDP Growth":"#3dcba8","Avg Geo Risk":"#e8a230","Top Overweight":"#2ec4a0","Top Underweight":"#e8537a","RF Accuracy":"#7c5cbf","Headlines Analysed":"#5e8fef"}
+    def card(icon,label,value,colour):
+        if isinstance(value,str) and len(value)>18: value=value[:16]+"…"
+        return (f'<div style="background:rgba(255,255,255,.72);border-radius:20px;padding:18px 14px 16px;text-align:center;border-top:3px solid {colour};box-shadow:0 4px 16px rgba(40,9,109,.08)">'
+                f'<div style="font-size:26px;margin-bottom:7px">{icon}</div>'
+                f'<div style="color:#9d8fc4;font-size:9.5px;text-transform:uppercase;letter-spacing:1.8px;margin-bottom:7px;font-weight:800">{label}</div>'
+                f'<div style="color:#2d1f4e;font-size:16px;font-weight:800">{value}</div></div>')
+    html='<div style="display:grid;grid-template-columns:repeat(auto-fit,minmax(130px,1fr));gap:12px;margin-bottom:24px">'
+    for key,val in kpis.items():
+        html+=card(icons.get(key,"📊"),key.replace("_"," ").title(),str(val),colours.get(key,"#8fa8f8"))
+    return html+"</div>"
+
+PAL=["#28096D","#2ec4a0","#e8537a","#F2C637","#5e8fef","#c45ea8"]
+def _styled(**kw):
+    d=dict(template="plotly_white",paper_bgcolor="rgba(255,255,255,0.95)",plot_bgcolor="rgba(255,255,255,0.98)",
+           font=dict(family="system-ui,sans-serif",color="#2d1f4e",size=12),margin=dict(l=60,r=20,t=70,b=70),
+           legend=dict(orientation="h",yanchor="bottom",y=1.02,xanchor="right",x=1),
+           title=dict(font=dict(size=15,color="#28096D")))
+    d.update(kw); return d
+def _empty_chart(title):
+    fig=go.Figure(); fig.update_layout(title=title,height=420,template="plotly_white",paper_bgcolor="rgba(255,255,255,0.95)",annotations=[dict(text="Run the pipeline to generate data",x=0.5,y=0.5,xref="paper",yref="paper",showarrow=False,font=dict(size=14,color="rgba(40,9,109,0.4)"))]); return fig
+def build_gdp_trend_chart():
+    p=PY_TAB_DIR/"df_dashboard.csv"
+    if not p.exists(): return _empty_chart("GDP Trend — run pipeline first")
+    df=pd.read_csv(p); yc=next((c for c in df.columns if "year" in c.lower()),df.columns[0]); vc=[c for c in df.columns if c!=yc and df[c].dtype in("float64","int64")]
+    if not vc: return _empty_chart("No numeric columns")
+    fig=go.Figure()
+    for i,col in enumerate(vc):
+        fig.add_trace(go.Scatter(x=df[yc],y=df[col],name=col.replace("_"," ").title(),mode="lines+markers",line=dict(color=PAL[i%len(PAL)],width=2.5),fill="tozeroy",fillcolor="rgba(40,9,109,0.06)"))
+    fig.update_layout(**_styled(height=450,hovermode="x unified",title=dict(text="Average EM GDP Growth Rate (2000–2023)"))); fig.update_yaxes(ticksuffix="%"); return fig
+def build_signal_chart():
+    p=PY_TAB_DIR/"investment_signal_summary.csv"
+    if not p.exists(): return _empty_chart("Investment Signal — run pipeline first")
+    df=pd.read_csv(p).sort_values("pct_overweight",ascending=True)
+    colors=["#2ec4a0" if v>=50 else "#e8537a" for v in df["pct_overweight"]]
+    fig=go.Figure(go.Bar(y=df["country"],x=df["pct_overweight"],orientation="h",marker_color=colors,hovertemplate="<b>%{y}</b><br>%{x:.1f}% years overweight<extra></extra>"))
+    fig.add_vline(x=50,line_dash="dash",line_color="gray",line_width=1)
+    fig.update_layout(**_styled(height=550,showlegend=False,title=dict(text="Investment Signal by Country (% Years Overweight)"))); fig.update_xaxes(title="% Years Overweight",ticksuffix="%"); return fig
+def build_sentiment_chart():
+    p=PY_TAB_DIR/"vader_by_country.csv"
+    if not p.exists(): return _empty_chart("Sentiment — run pipeline first")
+    df=pd.read_csv(p).sort_values("avg_vader_score")
+    colors=["#2ec4a0" if v>=0.05 else("#e8537a" if v<=-0.05 else "#5e8fef") for v in df["avg_vader_score"]]
+    fig=go.Figure(go.Bar(y=df["country"],x=df["avg_vader_score"],orientation="h",marker_color=colors,hovertemplate="<b>%{y}</b><br>VADER: %{x:.3f}<extra></extra>"))
+    fig.add_vline(x=0,line_dash="dot",line_color="gray")
+    fig.update_layout(**_styled(height=550,showlegend=False,title=dict(text="Average VADER News Sentiment by Country"))); fig.update_xaxes(title="Avg VADER Compound Score"); return fig
+def build_arima_chart():
+    hp=PY_TAB_DIR/"df_dashboard.csv"; fp=PY_TAB_DIR/"arima_gdp_forecast.csv"
+    if not fp.exists(): return _empty_chart("ARIMA Forecast — run pipeline first")
+    fc=pd.read_csv(fp); fig=go.Figure()
+    if hp.exists():
+        h=pd.read_csv(hp); yc=next((c for c in h.columns if "year" in c.lower()),h.columns[0]); vc=next((c for c in h.columns if "gdp" in c.lower() or "avg" in c.lower()),h.columns[-1])
+        fig.add_trace(go.Scatter(x=h[yc],y=h[vc],name="Historical",mode="lines+markers",line=dict(color="#28096D",width=2.5)))
+    yfc=next((c for c in fc.columns if "year" in c.lower()),fc.columns[0]); vfc=next((c for c in fc.columns if "forecast" in c.lower()),fc.columns[1])
+    fig.add_trace(go.Scatter(x=fc[yfc],y=fc[vfc],name="Forecast",mode="lines+markers",line=dict(color="#e8537a",width=2.5,dash="dash")))
+    if "lower_ci" in fc.columns and "upper_ci" in fc.columns:
+        fig.add_trace(go.Scatter(x=list(fc[yfc])+list(fc[yfc])[::-1],y=list(fc["upper_ci"])+list(fc["lower_ci"])[::-1],fill="toself",fillcolor="rgba(232,83,122,0.12)",line=dict(color="rgba(255,255,255,0)"),name="95% CI"))
+    fig.update_layout(**_styled(height=450,hovermode="x unified",title=dict(text="ARIMA GDP Growth Forecast (2024–2028)"))); fig.update_yaxes(ticksuffix="%"); return fig
+
+def refresh_dashboard(): return render_kpi_cards(),build_gdp_trend_chart(),build_signal_chart(),build_sentiment_chart()
+def refresh_gallery():
+    figs=[(str(p),p.stem.replace("_"," ").title()) for p in sorted(PY_FIG_DIR.glob("*.png"))]
+    idx=artifacts_index(); tc=idx["python"]["tables"]
+    df=_load_table_safe(PY_TAB_DIR/tc[0]) if tc else pd.DataFrame()
+    return (figs or []),gr.update(choices=tc,value=tc[0] if tc else None),df
+def on_table_select(choice):
+    if not choice: return pd.DataFrame([{"hint":"Select a table."}])
+    p=PY_TAB_DIR/choice; return _load_table_safe(p) if p.exists() else pd.DataFrame([{"error":f"Not found: {choice}"}])
 
 ensure_dirs()
+_ai_status=("Connected to **n8n workflow**." if N8N_WEBHOOK_URL else "Powered by **Claude** (Anthropic API)." if ANTHROPIC_API_KEY else "**LLM active**." if LLM_ENABLED else "Using **keyword matching**.")
 
-def load_css() -> str:
-    css_path = BASE_DIR / "style.css"
-    return css_path.read_text(encoding="utf-8") if css_path.exists() else ""
+def load_css():
+    p=BASE_DIR/"style.css"; return p.read_text(encoding="utf-8") if p.exists() else ""
 
-
-with gr.Blocks(title="AIBDM 2026 Workshop App") as demo:
-
-    gr.Markdown(
-        "# SE21 App Template\n"
-        "*This is an app template for SE21 students*",
-        elem_id="escp_title",
-    )
-
-    # ===========================================================
-    # TAB 1 -- Pipeline Runner
-    # ===========================================================
+with gr.Blocks(title="EM Geopolitical Analytics — ESCP SE21", css=load_css()) as demo:
+    gr.Markdown("# 🌍 Emerging Market Geopolitical Analytics\n*How can an EM investment fund use news sentiment & macro forecasting to allocate during geopolitical stress?*\n**Team:** Amaryllis · Kuang · Logan · Tommaso · Achille — ESCP SE21",elem_id="escp_title")
     with gr.Tab("Pipeline Runner"):
-        gr.Markdown()
-
-        with gr.Row():
-            with gr.Column(scale=1):
-                btn_nb1 = gr.Button("Step 1: Data Creation", variant="secondary")
-            with gr.Column(scale=1):
-                btn_nb2 = gr.Button("Step 2: Python Analysis", variant="secondary")
-
+        gr.Markdown("**Step 1** fetches World Bank macro data (20 countries, 2000–2023) and generates synthetic geopolitical risk scores, news headlines, and fund flows.\n\n**Step 2** runs VADER sentiment, ARIMA GDP forecasting, and Random Forest investment signal classification.")
         with gr.Row():
-            btn_all = gr.Button("Run Full Pipeline (Both Steps)", variant="primary")
-
-        run_log = gr.Textbox(
-            label="Execution Log",
-            lines=18,
-            max_lines=30,
-            interactive=False,
-        )
-
-        btn_nb1.click(run_datacreation, outputs=[run_log])
-        btn_nb2.click(run_pythonanalysis, outputs=[run_log])
-        btn_all.click(run_full_pipeline, outputs=[run_log])
-
-    # ===========================================================
-    # TAB 2 -- Dashboard (KPIs + Interactive Charts + Gallery)
-    # ===========================================================
+            btn1=gr.Button("▶ Step 1: Data Collection",variant="secondary"); btn2=gr.Button("▶ Step 2: Analysis",variant="secondary")
+        btn_all=gr.Button("⚡ Run Full Pipeline",variant="primary")
+        log=gr.Textbox(label="Execution Log",lines=22,max_lines=40,interactive=False)
+        btn1.click(run_datacreation,outputs=[log]); btn2.click(run_pythonanalysis,outputs=[log]); btn_all.click(run_full_pipeline,outputs=[log])
     with gr.Tab("Dashboard"):
-        kpi_html = gr.HTML(value=render_kpi_cards)
-
-        refresh_btn = gr.Button("Refresh Dashboard", variant="primary")
-
-        gr.Markdown("#### Interactive Charts")
-        chart_sales = gr.Plot(label="Monthly Overview")
-        chart_sentiment = gr.Plot(label="Sentiment Distribution")
-        chart_top = gr.Plot(label="Top Sellers")
-
-        gr.Markdown("#### Static Figures (from notebooks)")
-        gallery = gr.Gallery(
-            label="Generated Figures",
-            columns=2,
-            height=480,
-            object_fit="contain",
-        )
-
-        gr.Markdown("#### Data Tables")
-        table_dropdown = gr.Dropdown(
-            label="Select a table to view",
-            choices=[],
-            interactive=True,
-        )
-        table_display = gr.Dataframe(
-            label="Table Preview",
-            interactive=False,
-        )
-
-        def _on_refresh():
-            kpi, c1, c2, c3 = refresh_dashboard()
-            figs, dd, df = refresh_gallery()
-            return kpi, c1, c2, c3, figs, dd, df
-
-        refresh_btn.click(
-            _on_refresh,
-            outputs=[kpi_html, chart_sales, chart_sentiment, chart_top,
-                     gallery, table_dropdown, table_display],
-        )
-        table_dropdown.change(
-            on_table_select,
-            inputs=[table_dropdown],
-            outputs=[table_display],
-        )
-
-    # ===========================================================
-    # TAB 3 -- AI Dashboard
-    # ===========================================================
+        kpi_html=gr.HTML(value=render_kpi_cards); refresh_btn=gr.Button("🔄 Refresh Dashboard",variant="primary")
+        gr.Markdown("#### 📈 Interactive Charts")
+        chart_gdp=gr.Plot(label="GDP Growth Trend"); chart_signal=gr.Plot(label="Investment Signals"); chart_sent=gr.Plot(label="News Sentiment")
+        gr.Markdown("#### 🔮 ARIMA Forecast"); chart_arima=gr.Plot(label="GDP Forecast 2024–2028")
+        gr.Markdown("#### 🖼️ Figures"); gallery=gr.Gallery(label="Figures",columns=2,height=520)
+        gr.Markdown("#### 📋 Tables"); tbl_dd=gr.Dropdown(label="Select table",choices=[],interactive=True); tbl_df=gr.Dataframe(label="Preview",interactive=False)
+        def _refresh():
+            kpi,c1,c2,c3=refresh_dashboard(); c4=build_arima_chart(); figs,dd,df=refresh_gallery(); return kpi,c1,c2,c3,c4,figs,dd,df
+        refresh_btn.click(_refresh,outputs=[kpi_html,chart_gdp,chart_signal,chart_sent,chart_arima,gallery,tbl_dd,tbl_df])
+        tbl_dd.change(on_table_select,inputs=[tbl_dd],outputs=[tbl_df])
     with gr.Tab('"AI" Dashboard'):
-        _ai_status = (
-            "Connected to your **n8n workflow**." if N8N_WEBHOOK_URL
-            else "**LLM active.**" if LLM_ENABLED
-            else "Using **keyword matching**. Upgrade options: "
-                 "set `N8N_WEBHOOK_URL` to connect your n8n workflow, "
-                 "or set `HF_API_KEY` for direct LLM access."
-        )
-        gr.Markdown(
-            "### Ask questions, get interactive visualisations\n\n"
-            f"Type a question and the system will pick the right interactive chart or table. {_ai_status}"
-        )
-
-        with gr.Row(equal_height=True):
+        gr.Markdown(f"### 💬 Ask questions, get interactive visualisations\n\nAsk about country allocation, macro trends, sentiment, or forecasts. {_ai_status}")
+        with gr.Row():
             with gr.Column(scale=1):
-                chatbot = gr.Chatbot(
-                    label="Conversation",
-                    height=380,
-                )
-                user_input = gr.Textbox(
-                    label="Ask about your data",
-                    placeholder="e.g. Show me sales trends / What are the top sellers? / Sentiment analysis",
-                    lines=1,
-                )
-                gr.Examples(
-                    examples=[
-                        "Show me the sales trends",
-                        "What does the sentiment look like?",
-                        "Which titles sell the most?",
-                        "Show the ARIMA forecasts",
-                        "What are the pricing decisions?",
-                        "Give me a dashboard overview",
-                    ],
-                    inputs=user_input,
-                )
-
+                chatbot=gr.Chatbot(label="Conversation",height=420,type="messages")
+                user_input=gr.Textbox(label="Ask about your data",placeholder="e.g. Which countries to overweight? / ARIMA forecast / Turkey sentiment")
+                gr.Examples(examples=["Which countries should we overweight?","Show me the geopolitical risk heatmap","What does the ARIMA forecast say?","Which countries have negative news sentiment?","What features drive the investment signal?","Give me a dashboard overview"],inputs=user_input)
             with gr.Column(scale=1):
-                ai_figure = gr.Plot(
-                    label="Interactive Chart",
-                )
-                ai_table = gr.Dataframe(
-                    label="Data Table",
-                    interactive=False,
-                )
-
-        user_input.submit(
-            ai_chat,
-            inputs=[user_input, chatbot],
-            outputs=[chatbot, user_input, ai_figure, ai_table],
-        )
-
+                ai_chart=gr.Plot(label="Interactive Chart"); ai_table=gr.Dataframe(label="Data Table",interactive=False)
+        user_input.submit(ai_chat,inputs=[user_input,chatbot],outputs=[chatbot,user_input,ai_chart,ai_table])
 
-demo.launch(css=load_css(), allowed_paths=[str(BASE_DIR)])
+demo.launch(allowed_paths=[str(BASE_DIR)])

pythonanalysis.ipynb

@@ -0,0 +1,955 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "ce179b8c",
+   "metadata": {},
+   "source": [
+    "# 🌍 Notebook 2 — Data Analysis & Visualization\n",
+    "**Project:** EM Portfolio Risk Advisor\n",
+    "\n",
+    "**Research Question:** *How can an emerging market investment fund use news sentiment analysis and macroeconomic forecasting to identify which countries to overweight or underweight during periods of geopolitical stress?*\n",
+    "\n",
+    "**Team:** Amaryllis (PM) · Kuang (UX) · Tommaso (Data Analyst) · Logan (UX) · Achille (Content)\n",
+    "**Course:** AI for Big Data Management — ESCP SE21\n",
+    "\n",
+    "---\n",
+    "**Analyses performed in this notebook:**\n",
+    "| # | Type | Method |\n",
+    "|---|------|--------|\n",
+    "| 1 | Qualitative | VADER sentiment scoring of synthetic analyst reports |\n",
+    "| 2 | Quantitative | GDP growth heatmap (2000–2023) |\n",
+    "| 3 | Quantitative | Geopolitical risk heatmap |\n",
+    "| 4 | Quantitative | FDI & Inflation trend analysis |\n",
+    "| 5 | Quantitative | Random Forest investment signal classifier |\n",
+    "| 6 | Quantitative | ARIMA GDP growth forecasting (2024–2028) |\n",
+    "| 7 | Mixed | Country-level investment signal + sentiment composite |"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "fd142dab",
+   "metadata": {},
+   "source": [
+    "## 1. Install & Import"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "47a823f2",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "!pip install -q pandas numpy matplotlib seaborn vaderSentiment statsmodels scikit-learn plotly"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "0717b85c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import pandas as pd\n",
+    "import numpy as np\n",
+    "import matplotlib.pyplot as plt\n",
+    "import matplotlib.colors as mcolors\n",
+    "import seaborn as sns\n",
+    "import warnings, os, json\n",
+    "from pathlib import Path\n",
+    "from itertools import product\n",
+    "\n",
+    "warnings.filterwarnings('ignore')\n",
+    "np.random.seed(42)\n",
+    "\n",
+    "print('✅ Packages loaded')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "2aa18937",
+   "metadata": {},
+   "source": [
+    "## 2. Output Directory Setup\n",
+    "\n",
+    "All figures and tables are saved to `artifacts/` so the Hugging Face app can load them."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "89cb932e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "BASE_DIR  = Path('.')\n",
+    "ART_DIR   = BASE_DIR / 'artifacts'\n",
+    "PY_FIG    = ART_DIR / 'py' / 'figures'\n",
+    "PY_TAB    = ART_DIR / 'py' / 'tables'\n",
+    "\n",
+    "for p in [PY_FIG, PY_TAB]:\n",
+    "    p.mkdir(parents=True, exist_ok=True)\n",
+    "\n",
+    "print('✅ Output folders:')\n",
+    "print(' -', PY_FIG.resolve())\n",
+    "print(' -', PY_TAB.resolve())"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "59fccaec",
+   "metadata": {},
+   "source": [
+    "## 3. Load Datasets\n",
+    "\n",
+    "All five files were produced by Notebook 1 under the data contract."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ea5ed4b3",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "df_macro     = pd.read_csv('world_bank_macro.csv')\n",
+    "df_risk      = pd.read_csv('synthetic_risk_scores.csv')\n",
+    "df_sentiment = pd.read_csv('synthetic_news_sentiment.csv')\n",
+    "df_master    = pd.read_csv('title_level_features.csv')\n",
+    "df_monthly   = pd.read_csv('monthly_gdp_series.csv')\n",
+    "\n",
+    "print('world_bank_macro      :', df_macro.shape)\n",
+    "print('synthetic_risk_scores :', df_risk.shape)\n",
+    "print('synthetic_news_sent.  :', df_sentiment.shape)\n",
+    "print('title_level_features  :', df_master.shape)\n",
+    "print('monthly_gdp_series    :', df_monthly.shape)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "57b70ed2",
+   "metadata": {},
+   "source": [
+    "## 4. Data Quality Check"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "90a068f1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def quality_check(df, name):\n",
+    "    print(f'\\n🔍 {name}')\n",
+    "    print(f'  Shape : {df.shape}')\n",
+    "    nulls = df.isnull().sum()\n",
+    "    nulls = nulls[nulls > 0]\n",
+    "    print(f'  Nulls : {dict(nulls) if len(nulls) else \"none\"}')\n",
+    "    print(f'  Dtypes: {dict(df.dtypes)}')\n",
+    "    return df\n",
+    "\n",
+    "for df, nm in [\n",
+    "    (df_macro,     'world_bank_macro'),\n",
+    "    (df_risk,      'synthetic_risk_scores'),\n",
+    "    (df_sentiment, 'synthetic_news_sentiment'),\n",
+    "    (df_master,    'title_level_features'),\n",
+    "    (df_monthly,   'monthly_gdp_series'),\n",
+    "]:\n",
+    "    quality_check(df, nm)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "3d974373",
+   "metadata": {},
+   "source": [
+    "## 5. Qualitative Analysis — VADER Sentiment Scoring\n",
+    "\n",
+    "We apply the **VADER** (Valence Aware Dictionary and sEntiment Reasoner) lexicon to every\n",
+    "synthetic analyst report headline.  VADER returns a compound score in **[−1, +1]**:\n",
+    "\n",
+    "| Range | Label |\n",
+    "|-------|-------|\n",
+    "| ≥ 0.05 | bullish |\n",
+    "| ≤ −0.05 | bearish |\n",
+    "| otherwise | neutral |\n",
+    "\n",
+    "This mirrors how a portfolio analyst would extract sentiment from news feeds programmatically."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a0537f27",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer\n",
+    "\n",
+    "analyzer = SentimentIntensityAnalyzer()\n",
+    "\n",
+    "def vader_compound(text):\n",
+    "    return analyzer.polarity_scores(str(text))['compound']\n",
+    "\n",
+    "def vader_label(score):\n",
+    "    if score >= 0.05:  return 'bullish'\n",
+    "    if score <= -0.05: return 'bearish'\n",
+    "    return 'neutral'\n",
+    "\n",
+    "df_sentiment['vader_compound'] = df_sentiment['report_text'].apply(vader_compound)\n",
+    "df_sentiment['vader_label']    = df_sentiment['vader_compound'].apply(vader_label)\n",
+    "\n",
+    "print('VADER label distribution:')\n",
+    "print(df_sentiment['vader_label'].value_counts())\n",
+    "print()\n",
+    "print(df_sentiment[['country','year','sentiment_label','vader_compound','vader_label']].head(10))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1822af7e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Aggregate VADER scores per country\n",
+    "vader_by_country = (\n",
+    "    df_sentiment\n",
+    "    .groupby(['iso3','country'])\n",
+    "    .agg(avg_vader_score=('vader_compound','mean'),\n",
+    "         report_count=('vader_compound','count'))\n",
+    "    .reset_index()\n",
+    "    .sort_values('avg_vader_score')\n",
+    ")\n",
+    "vader_by_country.to_csv(PY_TAB / 'vader_by_country.csv', index=False)\n",
+    "print('✅ vader_by_country.csv saved')\n",
+    "print(vader_by_country)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "902ea1c0",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Aggregate VADER scores per country-year (for merging into master)\n",
+    "vader_agg = (\n",
+    "    df_sentiment\n",
+    "    .groupby(['iso3','year'])['vader_compound']\n",
+    "    .mean()\n",
+    "    .reset_index()\n",
+    "    .rename(columns={'vader_compound': 'vader_score'})\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "3ae2b5ee",
+   "metadata": {},
+   "source": [
+    "## 6. Merge VADER Scores into Master Dataset"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4bbfc012",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Also merge risk scores into master (if not already present)\n",
+    "risk_cols = ['iso3','year','geopolitical_risk_score']\n",
+    "if 'geopolitical_risk_score' not in df_master.columns:\n",
+    "    df_master = df_master.merge(\n",
+    "        df_risk[risk_cols], on=['iso3','year'], how='left'\n",
+    "    )\n",
+    "\n",
+    "# Merge VADER\n",
+    "if 'vader_score' not in df_master.columns:\n",
+    "    df_master = df_master.merge(vader_agg, on=['iso3','year'], how='left')\n",
+    "\n",
+    "# Clean numerics\n",
+    "for col in ['gdp_growth','fdi_pct_gdp','inflation','geopolitical_risk_score','vader_score']:\n",
+    "    if col in df_master.columns:\n",
+    "        df_master[col] = pd.to_numeric(df_master[col], errors='coerce')\n",
+    "\n",
+    "print('Master dataset shape after merge:', df_master.shape)\n",
+    "print(df_master[['country','year','gdp_growth','geopolitical_risk_score','vader_score']].head(10))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "6f506667",
+   "metadata": {},
+   "source": [
+    "## 7. Quantitative Analysis — GDP Growth Heatmap\n",
+    "\n",
+    "A heatmap gives fund managers an at-a-glance view of **which countries experienced growth\n",
+    "shocks vs. booms** over the 2000–2023 period.  Red cells correspond to known crises\n",
+    "(Argentina 2001–02, global GFC 2009, COVID 2020)."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d33ff313",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "PALETTE = ['#28096D','#2ec4a0','#e8537a','#F2C637','#5e8fef',\n",
+    "           '#c45ea8','#3dbacc','#a0522d','#6aaa3a','#d46060']\n",
+    "\n",
+    "# Build pivot\n",
+    "pivot_gdp = df_macro.pivot_table(index='country', columns='year', values='gdp_growth')\n",
+    "\n",
+    "fig, ax = plt.subplots(figsize=(18, 7))\n",
+    "sns.heatmap(\n",
+    "    pivot_gdp,\n",
+    "    cmap='RdYlGn',\n",
+    "    center=0,\n",
+    "    linewidths=0.4,\n",
+    "    linecolor='white',\n",
+    "    annot=False,\n",
+    "    fmt='.1f',\n",
+    "    ax=ax,\n",
+    "    cbar_kws={'label': 'GDP Growth (%)', 'shrink': 0.8}\n",
+    ")\n",
+    "ax.set_title('EM GDP Growth Rate (%) — 2000 to 2023', fontsize=15,\n",
+    "             fontweight='bold', color='#28096D', pad=16)\n",
+    "ax.set_xlabel('Year', fontsize=11, color='#28096D')\n",
+    "ax.set_ylabel('Country', fontsize=11, color='#28096D')\n",
+    "ax.tick_params(axis='x', labelsize=8, rotation=45)\n",
+    "ax.tick_params(axis='y', labelsize=10)\n",
+    "plt.tight_layout()\n",
+    "plt.savefig(PY_FIG / 'gdp_heatmap.png', dpi=150, bbox_inches='tight')\n",
+    "plt.show()\n",
+    "print('✅ gdp_heatmap.png saved')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "01b2971c",
+   "metadata": {},
+   "source": [
+    "## 8. Quantitative Analysis — Geopolitical Risk Heatmap"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c9dc1711",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "pivot_risk = df_risk.pivot_table(\n",
+    "    index='country', columns='year', values='geopolitical_risk_score'\n",
+    ")\n",
+    "\n",
+    "fig, ax = plt.subplots(figsize=(18, 7))\n",
+    "sns.heatmap(\n",
+    "    pivot_risk,\n",
+    "    cmap='YlOrRd',\n",
+    "    linewidths=0.4,\n",
+    "    linecolor='white',\n",
+    "    annot=False,\n",
+    "    ax=ax,\n",
+    "    cbar_kws={'label': 'Risk Score (0–10)', 'shrink': 0.8}\n",
+    ")\n",
+    "ax.set_title('Synthetic Geopolitical Risk Score — 2000 to 2023', fontsize=15,\n",
+    "             fontweight='bold', color='#28096D', pad=16)\n",
+    "ax.set_xlabel('Year', fontsize=11, color='#28096D')\n",
+    "ax.set_ylabel('Country', fontsize=11, color='#28096D')\n",
+    "ax.tick_params(axis='x', labelsize=8, rotation=45)\n",
+    "ax.tick_params(axis='y', labelsize=10)\n",
+    "plt.tight_layout()\n",
+    "plt.savefig(PY_FIG / 'geo_risk_heatmap.png', dpi=150, bbox_inches='tight')\n",
+    "plt.show()\n",
+    "print('✅ geo_risk_heatmap.png saved')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "30066fb5",
+   "metadata": {},
+   "source": [
+    "## 9. Qualitative Analysis — VADER Sentiment by Country\n",
+    "\n",
+    "We compare the average VADER compound score to the ground-truth `sentiment_label` assigned during synthetic data generation to validate alignment."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "03c36db4",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "fig, axes = plt.subplots(1, 2, figsize=(16, 6))\n",
+    "\n",
+    "# Left: VADER compound score (bar)\n",
+    "colors = ['#2ec4a0' if v >= 0.05 else ('#e8537a' if v <= -0.05 else '#5e8fef')\n",
+    "          for v in vader_by_country['avg_vader_score']]\n",
+    "axes[0].barh(vader_by_country['country'], vader_by_country['avg_vader_score'],\n",
+    "             color=colors, edgecolor='white', linewidth=0.6)\n",
+    "axes[0].axvline(0, color='gray', linewidth=0.8, linestyle='--')\n",
+    "axes[0].set_title('Average VADER Compound Score by Country', fontweight='bold', color='#28096D')\n",
+    "axes[0].set_xlabel('Compound Score (−1 to +1)')\n",
+    "axes[0].tick_params(labelsize=10)\n",
+    "\n",
+    "# Right: Sentiment distribution stacked bar\n",
+    "sent_dist = (\n",
+    "    df_sentiment.groupby(['country','vader_label'])\n",
+    "    .size().unstack(fill_value=0)\n",
+    ")\n",
+    "for lbl in ['bullish','neutral','bearish']:\n",
+    "    if lbl not in sent_dist.columns:\n",
+    "        sent_dist[lbl] = 0\n",
+    "sent_dist = sent_dist.reindex(columns=['bullish','neutral','bearish'])\n",
+    "sent_colors = {'bullish': '#2ec4a0', 'neutral': '#5e8fef', 'bearish': '#e8537a'}\n",
+    "sent_dist.plot(kind='barh', stacked=True, ax=axes[1],\n",
+    "               color=[sent_colors[c] for c in sent_dist.columns],\n",
+    "               edgecolor='white', linewidth=0.4)\n",
+    "axes[1].set_title('Report Sentiment Distribution by Country', fontweight='bold', color='#28096D')\n",
+    "axes[1].set_xlabel('Number of Reports')\n",
+    "axes[1].legend(title='VADER Label', loc='lower right')\n",
+    "axes[1].tick_params(labelsize=10)\n",
+    "\n",
+    "plt.suptitle('Qualitative Analysis: News Sentiment Across Emerging Markets',\n",
+    "             fontsize=13, fontweight='bold', color='#28096D', y=1.02)\n",
+    "plt.tight_layout()\n",
+    "plt.savefig(PY_FIG / 'vader_sentiment.png', dpi=150, bbox_inches='tight')\n",
+    "plt.show()\n",
+    "print('✅ vader_sentiment.png saved')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "67044680",
+   "metadata": {},
+   "source": [
+    "## 10. Quantitative Analysis — FDI & Inflation Trends"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c624d67d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "fig, axes = plt.subplots(1, 2, figsize=(16, 6))\n",
+    "\n",
+    "# FDI line chart\n",
+    "for i, (iso3, cname) in enumerate(df_macro.groupby(['iso3','country']).size().index):\n",
+    "    sub = df_macro[df_macro['iso3'] == iso3]\n",
+    "    axes[0].plot(sub['year'], sub['fdi_pct_gdp'],\n",
+    "                 label=cname, color=PALETTE[i % len(PALETTE)],\n",
+    "                 linewidth=1.8, alpha=0.85)\n",
+    "axes[0].set_title('FDI Inflows (% of GDP)', fontweight='bold', color='#28096D')\n",
+    "axes[0].set_xlabel('Year'); axes[0].set_ylabel('% of GDP')\n",
+    "axes[0].legend(fontsize=7, ncol=2)\n",
+    "axes[0].axhline(0, color='gray', linewidth=0.6, linestyle='--')\n",
+    "\n",
+    "# Inflation scatter / boxplot by country\n",
+    "df_macro_clean = df_macro[df_macro['inflation'].abs() < 200]\n",
+    "axes[1].boxplot(\n",
+    "    [df_macro_clean[df_macro_clean['country'] == c]['inflation'].dropna()\n",
+    "     for c in df_macro_clean['country'].unique()],\n",
+    "    labels=df_macro_clean['country'].unique(),\n",
+    "    patch_artist=True,\n",
+    "    boxprops=dict(facecolor='#a48de8', alpha=0.7),\n",
+    "    medianprops=dict(color='#28096D', linewidth=2)\n",
+    ")\n",
+    "axes[1].set_title('Inflation Distribution by Country', fontweight='bold', color='#28096D')\n",
+    "axes[1].set_ylabel('Inflation (%)')\n",
+    "axes[1].tick_params(axis='x', rotation=45, labelsize=9)\n",
+    "\n",
+    "plt.tight_layout()\n",
+    "plt.savefig(PY_FIG / 'fdi_inflation.png', dpi=150, bbox_inches='tight')\n",
+    "plt.show()\n",
+    "print('✅ fdi_inflation.png saved')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "69b4bbad",
+   "metadata": {},
+   "source": [
+    "## 11. Quantitative Analysis — Random Forest Investment Signal Classifier\n",
+    "\n",
+    "We train a **Random Forest** on macro + risk + sentiment features to learn which\n",
+    "country-years should be tagged *overweight*, *neutral*, or *underweight*.\n",
+    "\n",
+    "The investment signal target is defined by a rule:\n",
+    "- **Overweight**: GDP growth above median AND geopolitical risk below median\n",
+    "- **Underweight**: GDP growth in bottom tercile\n",
+    "- **Neutral**: all others\n",
+    "\n",
+    "This rule is used to create labelled training data.  The RF then learns non-linear\n",
+    "feature interactions that a simple rule would miss."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "cbec8543",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from sklearn.ensemble import RandomForestClassifier\n",
+    "from sklearn.model_selection import train_test_split\n",
+    "from sklearn.metrics import classification_report, accuracy_score\n",
+    "from sklearn.preprocessing import LabelEncoder\n",
+    "\n",
+    "FEATURES = ['gdp_growth', 'fdi_pct_gdp', 'inflation',\n",
+    "            'geopolitical_risk_score', 'vader_score']\n",
+    "\n",
+    "# Build target labels\n",
+    "df_ml = df_master.dropna(subset=[c for c in FEATURES if c in df_master.columns]).copy()\n",
+    "\n",
+    "gdp_med  = df_ml['gdp_growth'].median()\n",
+    "risk_med = df_ml['geopolitical_risk_score'].median() if 'geopolitical_risk_score' in df_ml.columns else 5\n",
+    "gdp_q33  = df_ml['gdp_growth'].quantile(0.33)\n",
+    "\n",
+    "def label_signal(row):\n",
+    "    try:\n",
+    "        if row['gdp_growth'] > gdp_med and row['geopolitical_risk_score'] < risk_med:\n",
+    "            return 'overweight'\n",
+    "        if row['gdp_growth'] < gdp_q33:\n",
+    "            return 'underweight'\n",
+    "        return 'neutral'\n",
+    "    except Exception:\n",
+    "        return 'neutral'\n",
+    "\n",
+    "df_ml['investment_signal'] = df_ml.apply(label_signal, axis=1)\n",
+    "print('Signal distribution:')\n",
+    "print(df_ml['investment_signal'].value_counts())"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "eb472edc",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Filter to available features\n",
+    "avail_features = [f for f in FEATURES if f in df_ml.columns]\n",
+    "X = df_ml[avail_features].fillna(df_ml[avail_features].median())\n",
+    "y = df_ml['investment_signal']\n",
+    "\n",
+    "X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2,\n",
+    "                                                     random_state=42, stratify=y)\n",
+    "\n",
+    "rf = RandomForestClassifier(n_estimators=200, max_depth=8,\n",
+    "                             class_weight='balanced', random_state=42)\n",
+    "rf.fit(X_train, y_train)\n",
+    "y_pred = rf.predict(X_test)\n",
+    "\n",
+    "print(classification_report(y_test, y_pred))\n",
+    "print(f'Accuracy: {accuracy_score(y_test, y_pred):.3f}')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "726e9575",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Feature importance chart\n",
+    "feat_imp = pd.DataFrame({\n",
+    "    'feature':    avail_features,\n",
+    "    'importance': rf.feature_importances_\n",
+    "}).sort_values('importance')\n",
+    "\n",
+    "fig, ax = plt.subplots(figsize=(9, 5))\n",
+    "colors_imp = ['#28096D' if f == feat_imp.iloc[-1]['feature'] else '#a48de8'\n",
+    "              for f in feat_imp['feature']]\n",
+    "ax.barh(feat_imp['feature'], feat_imp['importance'],\n",
+    "        color=colors_imp, edgecolor='white')\n",
+    "ax.set_title('Random Forest — Feature Importances', fontweight='bold',\n",
+    "             color='#28096D', fontsize=13)\n",
+    "ax.set_xlabel('Importance (Gini)', fontsize=11)\n",
+    "ax.tick_params(labelsize=11)\n",
+    "for v, f in zip(feat_imp['importance'], feat_imp['feature']):\n",
+    "    ax.text(v + 0.002, f, f'{v:.3f}', va='center', fontsize=10)\n",
+    "plt.tight_layout()\n",
+    "plt.savefig(PY_FIG / 'rf_feature_importance.png', dpi=150, bbox_inches='tight')\n",
+    "plt.show()\n",
+    "print('✅ rf_feature_importance.png saved')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "176186e1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Investment signal summary\n",
+    "df_ml['rf_prediction'] = rf.predict(X)\n",
+    "signal_summary = (\n",
+    "    df_ml.groupby('country')\n",
+    "    .apply(lambda x: (x['rf_prediction'] == 'overweight').mean() * 100)\n",
+    "    .reset_index()\n",
+    ")\n",
+    "signal_summary.columns = ['country', 'pct_overweight']\n",
+    "signal_summary = signal_summary.sort_values('pct_overweight', ascending=False)\n",
+    "signal_summary.to_csv(PY_TAB / 'investment_signal_summary.csv', index=False)\n",
+    "print('✅ investment_signal_summary.csv saved')\n",
+    "print(signal_summary)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "fed3baa1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Investment signal chart\n",
+    "fig, ax = plt.subplots(figsize=(10, 6))\n",
+    "colors_sig = ['#2ec4a0' if v >= 50 else '#e8537a'\n",
+    "              for v in signal_summary['pct_overweight']]\n",
+    "ax.barh(signal_summary['country'], signal_summary['pct_overweight'],\n",
+    "        color=colors_sig, edgecolor='white', linewidth=0.5)\n",
+    "ax.axvline(50, color='gray', linewidth=1.2, linestyle='--', label='50% threshold')\n",
+    "ax.set_title('Investment Signal: % of Years Classified as Overweight',\n",
+    "             fontweight='bold', color='#28096D', fontsize=13)\n",
+    "ax.set_xlabel('% Years Overweight', fontsize=11)\n",
+    "ax.set_xlim(0, 100)\n",
+    "for v, c in zip(signal_summary['pct_overweight'], signal_summary['country']):\n",
+    "    ax.text(v + 1, c, f'{v:.0f}%', va='center', fontsize=10)\n",
+    "ax.legend(fontsize=10)\n",
+    "plt.tight_layout()\n",
+    "plt.savefig(PY_FIG / 'investment_signal.png', dpi=150, bbox_inches='tight')\n",
+    "plt.show()\n",
+    "print('✅ investment_signal.png saved')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "a9005249",
+   "metadata": {},
+   "source": [
+    "## 12. Quantitative Analysis — ARIMA GDP Growth Forecasting\n",
+    "\n",
+    "We use **ARIMA** (Auto-Regressive Integrated Moving Average) to forecast the average\n",
+    "EM GDP growth rate for 2024–2028.  ARIMA is fit on the *average* monthly series across\n",
+    "all 10 countries, as this gives the fund a macro-level view of EM momentum.\n",
+    "\n",
+    "The `find_best_arima` helper searches across p ∈ [0,3], d ∈ [0,1], q ∈ [0,1] and\n",
+    "selects the order minimising **AIC**."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "44eff2cc",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from statsmodels.tsa.arima.model import ARIMA\n",
+    "\n",
+    "def find_best_arima(series, p_range=(0, 3), d_range=(0, 1), q_range=(0, 1)):\n",
+    "    best_aic, best_order, best_model = float('inf'), None, None\n",
+    "    for p, d, q in product(range(p_range[0], p_range[1] + 1),\n",
+    "                            range(d_range[0], d_range[1] + 1),\n",
+    "                            range(q_range[0], q_range[1] + 1)):\n",
+    "        try:\n",
+    "            m = ARIMA(series, order=(p, d, q)).fit()\n",
+    "            if m.aic < best_aic:\n",
+    "                best_aic, best_order, best_model = m.aic, (p, d, q), m\n",
+    "        except Exception:\n",
+    "            pass\n",
+    "    return best_order, best_model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6ace7aa0",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Prepare monthly series: average across countries\n",
+    "df_monthly['month'] = pd.to_datetime(df_monthly['month'], errors='coerce')\n",
+    "monthly_avg = (\n",
+    "    df_monthly.dropna(subset=['month'])\n",
+    "    .groupby('month')['gdp_growth_monthly']\n",
+    "    .mean()\n",
+    "    .sort_index()\n",
+    ")\n",
+    "\n",
+    "# Fit best ARIMA\n",
+    "best_order, best_fit = find_best_arima(monthly_avg)\n",
+    "print(f'Best ARIMA order: {best_order}')\n",
+    "print(f'AIC: {best_fit.aic:.2f}')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "9ab3a37c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Forecast 60 months ahead (5 years)\n",
+    "n_forecast = 60\n",
+    "forecast_result = best_fit.get_forecast(steps=n_forecast)\n",
+    "forecast_mean   = forecast_result.predicted_mean\n",
+    "forecast_ci     = forecast_result.conf_int()\n",
+    "\n",
+    "forecast_index = pd.date_range(\n",
+    "    start=monthly_avg.index[-1] + pd.DateOffset(months=1),\n",
+    "    periods=n_forecast, freq='MS'\n",
+    ")\n",
+    "forecast_mean.index = forecast_index\n",
+    "forecast_ci.index   = forecast_index\n",
+    "\n",
+    "# Annual roll-up for app\n",
+    "arima_annual = (\n",
+    "    pd.DataFrame({'forecast': forecast_mean,\n",
+    "                  'lower_ci': forecast_ci.iloc[:, 0],\n",
+    "                  'upper_ci': forecast_ci.iloc[:, 1]})\n",
+    "    .resample('YE').mean()\n",
+    ")\n",
+    "arima_annual.index = arima_annual.index.year\n",
+    "arima_annual.index.name = 'year'\n",
+    "arima_annual.reset_index().to_csv(PY_TAB / 'arima_gdp_forecast.csv', index=False)\n",
+    "print('✅ arima_gdp_forecast.csv saved')\n",
+    "print(arima_annual.round(3))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "3b8d4a55",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Plot historical + forecast\n",
+    "fig, ax = plt.subplots(figsize=(14, 6))\n",
+    "\n",
+    "# Historical (annual average for readability)\n",
+    "hist_annual = monthly_avg.resample('YE').mean()\n",
+    "hist_annual.index = hist_annual.index.year\n",
+    "ax.plot(hist_annual.index, hist_annual.values,\n",
+    "        color='#28096D', linewidth=2.5, marker='o', markersize=5, label='Historical avg')\n",
+    "\n",
+    "# Forecast\n",
+    "ax.plot(arima_annual.index, arima_annual['forecast'],\n",
+    "        color='#e8537a', linewidth=2.5, linestyle='--', marker='s', markersize=5, label='ARIMA Forecast')\n",
+    "ax.fill_between(arima_annual.index, arima_annual['lower_ci'], arima_annual['upper_ci'],\n",
+    "                color='#e8537a', alpha=0.15, label='95% CI')\n",
+    "\n",
+    "ax.axvline(2023.5, color='gray', linewidth=1, linestyle=':', label='Forecast start')\n",
+    "ax.axhline(0, color='black', linewidth=0.7, linestyle='-', alpha=0.4)\n",
+    "ax.set_title('ARIMA GDP Growth Forecast — Average EM Basket (2024–2028)',\n",
+    "             fontweight='bold', color='#28096D', fontsize=13)\n",
+    "ax.set_xlabel('Year', fontsize=11); ax.set_ylabel('Avg GDP Growth (%)', fontsize=11)\n",
+    "ax.yaxis.set_major_formatter(plt.FuncFormatter(lambda v, _: f'{v:.1f}%'))\n",
+    "ax.legend(fontsize=10)\n",
+    "plt.tight_layout()\n",
+    "plt.savefig(PY_FIG / 'arima_gdp_forecast.png', dpi=150, bbox_inches='tight')\n",
+    "plt.show()\n",
+    "print('✅ arima_gdp_forecast.png saved')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "6856f8f5",
+   "metadata": {},
+   "source": [
+    "## 13. Country Predictions Summary Table\n",
+    "\n",
+    "Final table merging the Random Forest signal, VADER score, and latest macro values into one actionable view for fund managers."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "07aac448",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Latest year for each country\n",
+    "latest = df_ml.sort_values('year').groupby('country').last().reset_index()\n",
+    "\n",
+    "latest_out = latest[['country', 'iso3', 'year',\n",
+    "                      'gdp_growth', 'fdi_pct_gdp', 'inflation',\n",
+    "                      'geopolitical_risk_score', 'vader_score',\n",
+    "                      'investment_signal', 'rf_prediction']].copy()\n",
+    "\n",
+    "# Readable labels\n",
+    "latest_out['recommendation'] = latest_out['rf_prediction'].map({\n",
+    "    'overweight':  'OVERWEIGHT  ✅',\n",
+    "    'underweight': 'UNDERWEIGHT 🔻',\n",
+    "    'neutral':     'NEUTRAL  ➡️'\n",
+    "})\n",
+    "\n",
+    "latest_out.to_csv(PY_TAB / 'country_predictions_latest.csv', index=False)\n",
+    "print('✅ country_predictions_latest.csv saved')\n",
+    "print(latest_out[['country','gdp_growth','geopolitical_risk_score',\n",
+    "                   'vader_score','recommendation']].to_string(index=False))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "0a01cf38",
+   "metadata": {},
+   "source": [
+    "## 14. Dashboard Data Exports\n",
+    "\n",
+    "The Hugging Face app (`app.py`) reads `df_dashboard.csv` and `kpis.json`."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1999318c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# df_dashboard: annual average GDP growth (used by the GDP trend chart in the app)\n",
+    "df_dashboard = (\n",
+    "    df_macro.groupby('year')['gdp_growth']\n",
+    "    .mean()\n",
+    "    .reset_index()\n",
+    "    .rename(columns={'gdp_growth': 'avg_gdp_growth'})\n",
+    ")\n",
+    "df_dashboard.to_csv(PY_TAB / 'df_dashboard.csv', index=False)\n",
+    "print('✅ df_dashboard.csv saved')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "fd440bbd",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# KPI summary\n",
+    "top_ow = signal_summary.iloc[0]['country']\n",
+    "top_uw = signal_summary.iloc[-1]['country']\n",
+    "\n",
+    "kpis = {\n",
+    "    'Countries Analysed': int(df_macro['country'].nunique()),\n",
+    "    'Years Covered':      f\"{int(df_macro['year'].min())}–{int(df_macro['year'].max())}\",\n",
+    "    'Avg GDP Growth':     f\"{df_macro['gdp_growth'].mean():.2f}%\",\n",
+    "    'Avg Geo Risk':       f\"{df_risk['geopolitical_risk_score'].mean():.1f}/10\",\n",
+    "    'Top Overweight':     str(top_ow),\n",
+    "    'Top Underweight':    str(top_uw),\n",
+    "    'RF Accuracy':        f\"{accuracy_score(y_test, y_pred)*100:.1f}%\",\n",
+    "    'Headlines Analysed': int(len(df_sentiment)),\n",
+    "}\n",
+    "\n",
+    "with open(PY_TAB / 'kpis.json', 'w') as f:\n",
+    "    json.dump(kpis, f, indent=2)\n",
+    "\n",
+    "print('✅ kpis.json saved')\n",
+    "print(json.dumps(kpis, indent=2))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "00a6bb3e",
+   "metadata": {},
+   "source": [
+    "## 15. Final Summary — 4-Panel Overview Chart"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "59a6cc07",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "fig, axes = plt.subplots(2, 2, figsize=(16, 12))\n",
+    "fig.suptitle('EM Portfolio Risk Advisor — Analysis Overview',\n",
+    "             fontsize=16, fontweight='bold', color='#28096D')\n",
+    "\n",
+    "# Panel A: Average GDP growth by country\n",
+    "avg_gdp = df_macro.groupby('country')['gdp_growth'].mean().sort_values()\n",
+    "clrs = ['#e8537a' if v < 0 else '#2ec4a0' for v in avg_gdp]\n",
+    "axes[0,0].barh(avg_gdp.index, avg_gdp.values, color=clrs, edgecolor='white')\n",
+    "axes[0,0].axvline(0, color='gray', linewidth=0.8, linestyle='--')\n",
+    "axes[0,0].set_title('A. Avg GDP Growth 2000–2023', fontweight='bold', color='#28096D')\n",
+    "axes[0,0].set_xlabel('% per year')\n",
+    "\n",
+    "# Panel B: Average geo risk by country\n",
+    "avg_risk = df_risk.groupby('country')['geopolitical_risk_score'].mean().sort_values()\n",
+    "axes[0,1].barh(avg_risk.index, avg_risk.values,\n",
+    "               color='#F2C637', edgecolor='white')\n",
+    "axes[0,1].set_title('B. Avg Geopolitical Risk Score', fontweight='bold', color='#28096D')\n",
+    "axes[0,1].set_xlabel('Score (0–10)')\n",
+    "\n",
+    "# Panel C: VADER compound score\n",
+    "axes[1,0].barh(vader_by_country['country'], vader_by_country['avg_vader_score'],\n",
+    "               color=['#2ec4a0' if v>=0.05 else('#e8537a' if v<=-0.05 else '#5e8fef')\n",
+    "                      for v in vader_by_country['avg_vader_score']],\n",
+    "               edgecolor='white')\n",
+    "axes[1,0].axvline(0, color='gray', linewidth=0.8, linestyle='--')\n",
+    "axes[1,0].set_title('C. Avg VADER Sentiment Score', fontweight='bold', color='#28096D')\n",
+    "axes[1,0].set_xlabel('Compound Score')\n",
+    "\n",
+    "# Panel D: % years overweight (RF signal)\n",
+    "sig_sorted = signal_summary.sort_values('pct_overweight')\n",
+    "axes[1,1].barh(sig_sorted['country'], sig_sorted['pct_overweight'],\n",
+    "               color=['#2ec4a0' if v>=50 else '#e8537a' for v in sig_sorted['pct_overweight']],\n",
+    "               edgecolor='white')\n",
+    "axes[1,1].axvline(50, color='gray', linewidth=1, linestyle='--')\n",
+    "axes[1,1].set_title('D. RF Investment Signal (% yrs Overweight)', fontweight='bold', color='#28096D')\n",
+    "axes[1,1].set_xlabel('% Years')\n",
+    "axes[1,1].set_xlim(0, 100)\n",
+    "\n",
+    "plt.tight_layout()\n",
+    "plt.savefig(PY_FIG / 'analysis_overview.png', dpi=150, bbox_inches='tight')\n",
+    "plt.show()\n",
+    "print('✅ analysis_overview.png saved')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "0a4810ef",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "print()\n",
+    "print('=' * 55)\n",
+    "print('  NOTEBOOK 2 COMPLETE — EM PORTFOLIO RISK ADVISOR')\n",
+    "print('=' * 55)\n",
+    "\n",
+    "import os\n",
+    "figures = list(PY_FIG.glob('*.png'))\n",
+    "tables  = list(PY_TAB.glob('*.csv')) + list(PY_TAB.glob('*.json'))\n",
+    "\n",
+    "print(f'  Figures saved : {len(figures)}')\n",
+    "for f in sorted(figures): print(f'    • {f.name}')\n",
+    "print(f'  Tables saved  : {len(tables)}')\n",
+    "for t in sorted(tables):  print(f'    • {t.name}')\n",
+    "print()\n",
+    "print('  Handoff to Hugging Face Space app.py ✅')"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "name": "python",
+   "version": "3.10.0"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

requirements.txt

@@ -1,17 +1,12 @@
-gradio==6.0.0
-pandas>=2.0.0
-numpy>=1.24.0
-matplotlib>=3.7.0
-seaborn>=0.13.0
-statsmodels>=0.14.0
-scikit-learn>=1.3.0
-papermill>=2.5.0
-nbformat>=5.9.0
-pillow>=10.0.0
-requests>=2.31.0
-beautifulsoup4>=4.12.0
-vaderSentiment>=3.3.2
-huggingface_hub>=0.20.0
-textblob>=0.18.0
-faker>=20.0.0
-plotly>=5.18.0
+gradio==5.25.0
+papermill
+pandas
+numpy
+matplotlib
+seaborn
+vaderSentiment
+statsmodels
+scikit-learn
+plotly
+wbgapi
+ipykernel

style.css

@@ -1,326 +1,276 @@
-/* --- Target the Gradio app wrapper for backgrounds --- */
-gradio-app,
-.gradio-app,
-.main,
-#app,
-[data-testid="app"] {
-  background-color: rgb(40,9,109) !important;
-  background-image:
-    url('https://huggingface.co/spaces/atascioglu/SE21AppTemplate/resolve/main/background_top.png'),
-    url('https://huggingface.co/spaces/atascioglu/SE21AppTemplate/resolve/main/background_mid.png'),
-    url('https://huggingface.co/spaces/atascioglu/SE21AppTemplate/resolve/main/background_bottom.png') !important;
-  background-position:
-    top center,
-    0 913px,
-    bottom center !important;
-  background-repeat:
-    no-repeat,
-    repeat-y,
-    no-repeat !important;
-  background-size:
-    100% auto,
-    100% auto,
-    100% auto !important;
-  min-height: 100vh !important;
-}
-
-/* --- Fallback on html/body --- */
-html, body {
-  background-color: rgb(40,9,109) !important;
-  margin: 0 !important;
-  padding: 0 !important;
-  min-height: 100vh !important;
-}
-
-/* Bottom image is now part of the main background layers (positioned at bottom center) */
-
-/* --- Main container --- */
+/* =========================================================
+   EM Portfolio Risk Advisor — Hugging Face Space Stylesheet
+   ESCP SE21 · Amaryllis · Kuang · Tommaso · Logan · Achille
+   ========================================================= */
+
+/* ── Root Variables ────────────────────────────────────── */
+:root {
+  --escp-deep:   #28096D;   /* ESCP signature purple */
+  --escp-mid:    #4a2a9e;
+  --escp-light:  #a48de8;
+  --escp-pale:   #ede8fc;
+  --teal:        #2ec4a0;
+  --coral:       #e8537a;
+  --gold:        #F2C637;
+  --blue:        #5e8fef;
+  --bg:          #f0ecfa;
+  --card-bg:     rgba(255,255,255,0.80);
+  --radius:      18px;
+  --shadow:      0 6px 28px rgba(40,9,109,0.12);
+}
+
+/* ── Page Background ───────────────────────────────────── */
+body,
 .gradio-container {
-  max-width: 1400px !important;
-  width: 94vw !important;
-  margin: 0 auto !important;
-  padding-top: 220px !important;
-  padding-bottom: 150px !important;
-  background: transparent !important;
+  background: linear-gradient(140deg, #1a0550 0%, #28096D 40%, #3d1a8a 70%, #1a0550 100%) !important;
+  min-height: 100vh;
+  font-family: "Inter", "Segoe UI", system-ui, -apple-system, sans-serif;
 }
 
-/* --- Title in ESCP gold --- */
-#escp_title h1 {
-  color: rgb(242,198,55) !important;
-  font-size: 3rem !important;
-  font-weight: 800 !important;
-  text-align: center !important;
-  margin: 0 0 12px 0 !important;
-}
-
-/* --- Subtitle --- */
-#escp_title p, #escp_title em {
-  color: rgba(255,255,255,0.85) !important;
-  text-align: center !important;
-}
-
-/* --- Tab bar background --- */
-.tabs > .tab-nav,
-.tab-nav,
-div[role="tablist"],
-.svelte-tabs > .tab-nav {
-  background: rgba(40,9,109,0.6) !important;
-  border-radius: 10px 10px 0 0 !important;
-  padding: 4px !important;
+/* ── Main Wrapper ──────────────────────────────────────── */
+.main.svelte-1kyws56,
+.wrap.svelte-1kyws56 {
+  background: transparent !important;
 }
 
-/* --- ALL tab buttons: force white text --- */
-.tabs > .tab-nav button,
-.tab-nav button,
-div[role="tablist"] button,
-button[role="tab"],
-.svelte-tabs button,
-.tab-nav > button,
-.tabs button {
-  color: #ffffff !important;
-  font-weight: 600 !important;
-  border: none !important;
-  background: transparent !important;
-  padding: 10px 20px !important;
-  border-radius: 8px 8px 0 0 !important;
-  opacity: 1 !important;
-}
-
-/* --- Selected tab: ESCP gold --- */
-.tabs > .tab-nav button.selected,
-.tab-nav button.selected,
-button[role="tab"][aria-selected="true"],
-button[role="tab"].selected,
-div[role="tablist"] button[aria-selected="true"],
-.svelte-tabs button.selected {
-  color: rgb(242,198,55) !important;
-  background: rgba(255,255,255,0.12) !important;
+/* ── Title / Header ────────────────────────────────────── */
+#escp_title {
+  background: rgba(255,255,255,0.08);
+  border: 1px solid rgba(255,255,255,0.18);
+  border-radius: var(--radius);
+  padding: 28px 32px 22px;
+  margin-bottom: 20px;
+  backdrop-filter: blur(10px);
 }
 
-/* --- Unselected tabs: ensure visibility --- */
-.tabs > .tab-nav button:not(.selected),
-.tab-nav button:not(.selected),
-button[role="tab"][aria-selected="false"],
-button[role="tab"]:not(.selected),
-div[role="tablist"] button:not([aria-selected="true"]) {
+#escp_title h1 {
   color: #ffffff !important;
-  opacity: 1 !important;
+  font-size: 1.8rem;
+  font-weight: 800;
+  margin-bottom: 6px;
 }
 
-/* --- White card panels --- */
-.gradio-container .gr-block,
-.gradio-container .gr-box,
-.gradio-container .gr-panel,
-.gradio-container .gr-group {
-  background: #ffffff !important;
-  border-radius: 10px !important;
+#escp_title p,
+#escp_title em {
+  color: rgba(255,255,255,0.80) !important;
+  font-size: 0.95rem;
 }
 
-/* --- Tab content area --- */
-.tabitem {
-  background: rgba(255,255,255,0.95) !important;
-  border-radius: 0 0 10px 10px !important;
-  padding: 16px !important;
+#escp_title strong {
+  color: var(--teal) !important;
 }
 
-/* --- Inputs --- */
-.gradio-container input,
-.gradio-container textarea,
-.gradio-container select {
-  background: #ffffff !important;
-  border: 1px solid #d1d5db !important;
-  border-radius: 8px !important;
+/* ── Tabs ──────────────────────────────────────────────── */
+.tabs > .tab-nav {
+  background: rgba(255,255,255,0.07) !important;
+  border-radius: 14px 14px 0 0 !important;
+  border-bottom: 1px solid rgba(255,255,255,0.14) !important;
+  padding: 6px 8px 0 !important;
 }
 
-/* --- Buttons: ESCP purple primary --- */
-.gradio-container button:not([role="tab"]) {
-  font-weight: 600 !important;
-  padding: 10px 16px !important;
-  border-radius: 10px !important;
+.tabs > .tab-nav button {
+  color: rgba(255,255,255,0.65) !important;
+  font-weight: 600;
+  font-size: 0.88rem;
+  padding: 9px 18px;
+  border-radius: 10px 10px 0 0 !important;
+  transition: all 0.2s ease;
 }
 
-button.primary {
-  background-color: rgb(40,9,109) !important;
+.tabs > .tab-nav button.selected {
+  background: rgba(255,255,255,0.18) !important;
   color: #ffffff !important;
-  border: none !important;
-}
-
-button.primary:hover {
-  background-color: rgb(60,20,140) !important;
+  border-bottom: 2px solid var(--teal) !important;
 }
 
-button.secondary {
-  background-color: #ffffff !important;
-  color: rgb(40,9,109) !important;
-  border: 2px solid rgb(40,9,109) !important;
-}
-
-button.secondary:hover {
-  background-color: rgb(240,238,250) !important;
+.tabs > .tab-nav button:hover:not(.selected) {
+  background: rgba(255,255,255,0.10) !important;
+  color: #ffffff !important;
 }
 
-/* --- Dataframes --- */
-[data-testid="dataframe"] {
-  background-color: #ffffff !important;
-  border-radius: 10px !important;
+/* ── Tab Content Panels ────────────────────────────────── */
+.tabitem {
+  background: rgba(255,255,255,0.06) !important;
+  border-radius: 0 0 var(--radius) var(--radius) !important;
+  border: 1px solid rgba(255,255,255,0.12) !important;
+  border-top: none !important;
+  padding: 24px !important;
+}
+
+/* ── Buttons ───────────────────────────────────────────── */
+button.primary,
+.gr-button-primary {
+  background: linear-gradient(135deg, var(--teal) 0%, #1da889 100%) !important;
+  color: #fff !important;
+  border: none !important;
+  border-radius: 12px !important;
+  font-weight: 700 !important;
+  padding: 11px 24px !important;
+  font-size: 0.92rem !important;
+  box-shadow: 0 4px 16px rgba(46,196,160,0.35) !important;
+  transition: transform 0.15s, box-shadow 0.15s !important;
 }
 
-table {
-  font-size: 0.85rem !important;
+button.primary:hover,
+.gr-button-primary:hover {
+  transform: translateY(-2px) !important;
+  box-shadow: 0 8px 24px rgba(46,196,160,0.45) !important;
 }
 
-/* --- Chatbot (AI Dashboard tab) --- */
-.gr-chatbot {
-  min-height: 380px !important;
-  background-color: #ffffff !important;
+button.secondary,
+.gr-button-secondary {
+  background: rgba(255,255,255,0.12) !important;
+  color: #fff !important;
+  border: 1px solid rgba(255,255,255,0.25) !important;
   border-radius: 12px !important;
+  font-weight: 600 !important;
+  padding: 11px 22px !important;
+  transition: background 0.15s !important;
 }
 
-.gr-chatbot .message.user {
-  background-color: rgb(232,225,250) !important;
-  border-radius: 12px !important;
+button.secondary:hover,
+.gr-button-secondary:hover {
+  background: rgba(255,255,255,0.22) !important;
 }
 
-.gr-chatbot .message.bot {
-  background-color: #f3f4f6 !important;
+/* ── Text Areas / Logs ─────────────────────────────────── */
+textarea,
+.scroll-hide {
+  background: rgba(10,3,30,0.55) !important;
+  color: #d4c8f8 !important;
+  border: 1px solid rgba(255,255,255,0.14) !important;
   border-radius: 12px !important;
+  font-family: "JetBrains Mono", "Fira Code", monospace !important;
+  font-size: 0.83rem !important;
 }
 
-/* --- Gallery --- */
-.gallery {
-  background: #ffffff !important;
+/* ── Textbox Inputs ────────────────────────────────────── */
+input[type="text"],
+.gr-text-input {
+  background: rgba(255,255,255,0.10) !important;
+  color: #fff !important;
+  border: 1px solid rgba(255,255,255,0.20) !important;
   border-radius: 10px !important;
+  padding: 10px 14px !important;
 }
 
-/* --- Log textbox --- */
-textarea {
-  font-family: monospace !important;
-  font-size: 0.8rem !important;
+input[type="text"]::placeholder {
+  color: rgba(255,255,255,0.40) !important;
 }
 
-/* --- Markdown headings inside tabs --- */
-.tabitem h3 {
-  color: rgb(40,9,109) !important;
-  font-weight: 700 !important;
+/* ── Labels & Markdown Text ────────────────────────────── */
+label, .label-wrap span {
+  color: rgba(255,255,255,0.75) !important;
+  font-size: 0.82rem !important;
+  font-weight: 600 !important;
+  text-transform: uppercase;
+  letter-spacing: 0.06em;
 }
 
-.tabitem h4 {
-  color: #374151 !important;
+.prose p, .prose li, .markdown-body p, .markdown-body li {
+  color: rgba(255,255,255,0.85) !important;
 }
 
-/* --- Examples row (AI Dashboard) --- */
-.examples-row button {
-  background: rgb(240,238,250) !important;
-  color: rgb(40,9,109) !important;
-  border: 1px solid rgb(40,9,109) !important;
-  border-radius: 8px !important;
-  font-size: 0.85rem !important;
+.prose h3, .prose h4, .markdown-body h3, .markdown-body h4 {
+  color: var(--teal) !important;
 }
 
-.examples-row button:hover {
-  background: rgb(232,225,250) !important;
+/* ── KPI Cards ─────────────────────────────────────────── */
+.kpi-card {
+  background: var(--card-bg);
+  border-radius: 18px;
+  padding: 18px 14px 16px;
+  text-align: center;
+  border-top: 3px solid var(--teal);
+  box-shadow: var(--shadow);
+  backdrop-filter: blur(8px);
+  transition: transform 0.2s;
 }
 
-/* --- Header / footer: transparent over banner --- */
-header, header *,
-footer, footer * {
-  background: transparent !important;
-  box-shadow: none !important;
-}
+.kpi-card:hover { transform: translateY(-3px); }
 
-footer a, footer button,
-header a, header button {
-  background: transparent !important;
-  border: none !important;
-  box-shadow: none !important;
+/* ── Chatbot ───────────────────────────────────────────── */
+.chatbot .message.user {
+  background: linear-gradient(135deg, var(--escp-mid), var(--escp-deep)) !important;
+  color: #fff !important;
+  border-radius: 16px 16px 4px 16px !important;
 }
 
-#footer, #footer *,
-[class*="footer"], [class*="footer"] *,
-[class*="chip"], [class*="pill"], [class*="chip"] *, [class*="pill"] * {
-  background: transparent !important;
-  border: none !important;
-  box-shadow: none !important;
+.chatbot .message.bot {
+  background: rgba(255,255,255,0.88) !important;
+  color: var(--escp-deep) !important;
+  border-radius: 4px 16px 16px 16px !important;
+  box-shadow: 0 2px 12px rgba(40,9,109,0.10) !important;
 }
 
-[data-testid*="api"], [data-testid*="settings"],
-[id*="api"], [id*="settings"],
-[class*="api"], [class*="settings"],
-[class*="bottom"], [class*="toolbar"], [class*="controls"] {
-  background: transparent !important;
-  box-shadow: none !important;
+/* ── Plots / Gallery ───────────────────────────────────── */
+.gr-plot, .plot-container {
+  background: rgba(255,255,255,0.95) !important;
+  border-radius: var(--radius) !important;
+  box-shadow: var(--shadow) !important;
+  padding: 4px !important;
 }
 
-[data-testid*="api"] *, [data-testid*="settings"] *,
-[id*="api"] *, [id*="settings"] *,
-[class*="api"] *, [class*="settings"] * {
-  background: transparent !important;
-  box-shadow: none !important;
+.gallery-item {
+  border-radius: 12px !important;
+  overflow: hidden !important;
+  box-shadow: var(--shadow) !important;
 }
 
-section footer {
-  background: transparent !important;
+/* ── Dataframe ─────────────────────────────────────────── */
+.gr-dataframe, table {
+  background: rgba(255,255,255,0.93) !important;
+  border-radius: var(--radius) !important;
+  overflow: hidden !important;
+  box-shadow: var(--shadow) !important;
 }
 
-section footer button,
-section footer a {
-  background: transparent !important;
-  background-color: transparent !important;
-  border: none !important;
-  box-shadow: none !important;
-  color: white !important;
+th {
+  background: var(--escp-deep) !important;
+  color: #fff !important;
+  font-weight: 700 !important;
+  text-transform: uppercase;
+  font-size: 0.78rem;
+  letter-spacing: 0.05em;
 }
 
-section footer button:hover,
-section footer button:focus,
-section footer a:hover,
-section footer a:focus {
-  background: transparent !important;
-  background-color: transparent !important;
-  box-shadow: none !important;
-}
+tr:nth-child(even) { background: var(--escp-pale) !important; }
+tr:hover { background: rgba(164,141,232,0.18) !important; }
 
-section footer button,
-section footer button * {
-  background: transparent !important;
-  background-color: transparent !important;
-  background-image: none !important;
-  box-shadow: none !important;
-  filter: none !important;
+/* ── Dropdown ────────────────���─────────────────────────── */
+.gr-dropdown select {
+  background: rgba(255,255,255,0.12) !important;
+  color: #fff !important;
+  border: 1px solid rgba(255,255,255,0.22) !important;
+  border-radius: 10px !important;
 }
 
-section footer button::before,
-section footer button::after {
-  background: transparent !important;
-  background-color: transparent !important;
-  background-image: none !important;
-  box-shadow: none !important;
-  filter: none !important;
+/* ── Scrollbars ────────────────────────────────────────── */
+::-webkit-scrollbar { width: 6px; height: 6px; }
+::-webkit-scrollbar-track { background: rgba(255,255,255,0.05); }
+::-webkit-scrollbar-thumb {
+  background: var(--escp-light);
+  border-radius: 3px;
 }
 
-section footer a,
-section footer a * {
-  background: transparent !important;
-  background-color: transparent !important;
-  box-shadow: none !important;
+/* ── Examples (chat suggestions) ──────────────────────── */
+.examples-list button {
+  background: rgba(255,255,255,0.08) !important;
+  color: rgba(255,255,255,0.80) !important;
+  border: 1px solid rgba(255,255,255,0.16) !important;
+  border-radius: 8px !important;
+  font-size: 0.82rem !important;
+  transition: background 0.15s !important;
 }
 
-.gradio-container footer button,
-.gradio-container footer button *,
-.gradio-container .footer button,
-.gradio-container .footer button * {
-  background: transparent !important;
-  background-color: transparent !important;
-  background-image: none !important;
-  box-shadow: none !important;
+.examples-list button:hover {
+  background: rgba(255,255,255,0.18) !important;
+  color: #fff !important;
 }
 
-.gradio-container footer button::before,
-.gradio-container footer button::after,
-.gradio-container .footer button::before,
-.gradio-container .footer button::after {
-  background: transparent !important;
-  background-color: transparent !important;
-  background-image: none !important;
-  box-shadow: none !important;
-}
+/* ── Responsive ────────────────────────────────────────── */
+@media (max-width: 768px) {
+  #escp_title { padding: 18px 16px; }
+  #escp_title h1 { font-size: 1.3rem; }
+  .tabitem { padding: 14px !important; }
+}