🏛️ fiscus Dashboard

An Executive-Grade Multi-Asset Portfolio Management & Intelligence System

🎯 Executive Summary (The Business Problem)

Family offices and private wealth managers struggle with a critical blind spot: the inability to view real-time, cross-currency asset performance in one centralized place. When capital is fragmented across illiquid private equity, volatile crypto, foreign real estate, and public equities, assessing true Net Worth and concentrated risk becomes a massive operational bottleneck.

Fiscus is built to solve this exact problem. It optimizes portfolio management by consolidating these disparate, multi-currency asset classes into a single, stakeholder-ready view. By automating data pipelines and normalizing FX rates in real-time, Fiscus eliminates the manual spreadsheet overhead. Furthermore, it leverages an ensemble of machine learning models (LSTM, Autoencoders, NLP) to filter raw market noise into actionable, confidence-rated investment signals, effectively upgrading traditional tracking into a proactive decision-support engine.

---

🔬 Research Features

Feature	Description
Walk-Forward Backtesting	Expanding-window evaluation with monthly rebalance steps over 2-3 years of data
Baseline Models	Naive, Random Walk, ARIMA (auto-order), Buy-and-Hold, SMA Crossover
Evaluation Metrics	RMSE, MAPE, Directional Accuracy, Hit Rate, Profit Factor, Sharpe, Calmar, Information Ratio
Statistical Tests	Paired t-test, Wilcoxon signed-rank, Diebold-Mariano test
Ablation Studies	15 configurations systematically removing components to quantify contribution
Sentiment-Weighted Ensemble	FinBERT score dynamically re-weights model ensemble via gating function
MC Dropout Uncertainty	50 stochastic forward passes produce confidence intervals that widen with horizon
F1-Calibrated Anomaly Detection	Threshold sweep (P80-P99) with synthetic labels for optimal anomaly sensitivity
Mathematical Formulations	All equations documented in LaTeX-exportable format

🏗️ Architecture & Data Flow

graph TD
    %% ETL Pipeline
    subgraph Data Layer: Automated ETL Pipeline
        A1[Raw Market Data<br>yfinance / FX Rates] --> B[Data Normalization]
        A2[Manual Entry / CSV Import<br>Illiquid Assets] --> B
        B --> C[Base Currency Conversion<br>Real-time FX]
        C --> D[(SQLAlchemy ORM<br>Portfolio Database)]
    end

    %% ML Pipeline
    subgraph Intelligence Layer: Multi-Model ML Pipeline
        D --> E{Feature Extraction}
        
        %% Models
        E --> F1[LSTM Neural Net<br>Non-linear Trends]
        E --> F2[Monte Carlo GBM<br>Stochastic Pathways]
        E --> F3[Holt's Smoothing<br>Exponential Trend]
        E --> F4[Autoencoder<br>Anomaly Detection]
        
        %% NLP
        G[Real-time News RSS] --> H[FinBERT Transformer<br>Sentiment Analysis]
        H --> I[Sentiment Gating Function]
        
        %% Fusion
        F1 & F2 & F3 --> J[Ensemble Orchestrator]
        I -. Modulates Weights .-> J
        J --> K[Confidence-Rated Signal<br>BUY / SELL / HOLD]
    end

    %% Presentation
    subgraph Presentation Layer
        D --> L1[Net Worth Aggregation]
        F4 --> L2[Risk Alerts]
        K --> L3[Strategic Advisory]
        
        L1 & L2 & L3 --> UI[Streamlit Dashboard<br>Stakeholder View]
    end

Loading

Directory Structure

Family_Office_Portfolio_Tracker/
├── app.py                    # Main Streamlit dashboard (entry point)
├── pages/
│   └── page_backtesting.py   # Walk-forward backtesting & ablation UI
│
├── # ── Core Data Layer ──
├── utils.py                  # Portfolio CRUD, calculations, XIRR
├── database.py               # SQLAlchemy ORM (SQLite)
├── market_data.py            # yfinance price fetcher + FX rates
├── target_allocation.py      # IPS engine & rebalancing
│
├── # ── Analytics ──
├── analytics_engine.py       # Risk metrics, Monte Carlo, Holt's smoothing
├── ai_advisor.py             # Data-driven advisory (portfolio-gap scoring)
│
├── # ── Deep Learning ──
├── dl_forecaster.py          # LSTM with MC Dropout uncertainty
├── dl_anomaly.py             # Autoencoder with F1-calibrated threshold
├── dl_sentiment.py           # FinBERT transformer sentiment analysis
├── intelligence_engine.py    # Ensemble orchestration + sentiment gating
│
├── # ── Research Infrastructure ──
├── backtesting.py            # Walk-forward backtesting engine
├── baselines.py              # Baseline models (Naive, RW, ARIMA, B&H, SMA)
├── evaluation.py             # Metrics suite + statistical significance tests
├── ablation_runner.py        # Automated ablation study executor
├── math_formulations.py      # LaTeX-exportable mathematical documentation
│
├── # ── Security ──
├── auth.py                   # streamlit-authenticator login wall
│
├── # ── Paper ──
├── paper/
│   └── paper_skeleton.tex    # LaTeX paper skeleton (IEEE format)
│
└── requirements.txt

🚀 Quick Start

# Install dependencies
pip install -r requirements.txt

# Launch dashboard
streamlit run app.py

Default credentials: admin / admin123

📊 Key Modules

Ensemble Forecasting with Sentiment Gating

The system runs three models simultaneously — LSTM, Monte Carlo GBM, and Holt's Exponential Smoothing — then fuses their predictions using inverse-MAPE weighting modulated by a FinBERT sentiment gating function:

w_i' = w_i^(base) × g(s, m_i)

g(s, LSTM)          = 1 + 0.3s
g(s, Monte Carlo)   = 1 - 0.4s  
g(s, Exp Smoothing) = 1 + 0.5s

Where s ∈ [-1, 1] is the aggregate FinBERT sentiment score. Bearish sentiment increases Monte Carlo weight (tail risk capture); bullish increases trend-following weights.

Walk-Forward Backtesting

from backtesting import WalkForwardBacktester

bt = WalkForwardBacktester("AAPL", lookback_years=3, forecast_horizon=21)
results = bt.run(models=['naive', 'arima', 'monte_carlo', 'exp_smoothing'])
summary = bt.get_summary()         # Per-model metrics
sig = bt.get_significance_tests()  # Pairwise statistical tests

Ablation Studies

from ablation_runner import AblationRunner

runner = AblationRunner(tickers=["AAPL", "MSFT", "NVDA"])
runner.run_suite(configs=["full_ensemble", "no_sentiment", "naive_baseline"])
comparison = runner.get_averaged_comparison()  # Averaged across tickers

📈 Sample Portfolio

The dashboard ships with a comprehensive sample portfolio spanning:

• 13 asset categories (Public Equity, Private Equity, Real Estate, Gold, Crypto, etc.)
• 6 currencies (USD, INR, EUR, GBP, JPY, AED)
• 55+ holdings with realistic valuations
• 11 liability positions (mortgages, loans, credit cards)

🔐 Security

• Authentication: streamlit-authenticator with cookie-based sessions
• Data Storage: SQLAlchemy ORM with SQLite (local, encrypted at rest via OS-level encryption)
• Input Validation: Schema validation on CSV imports

📄 Citation

@software{family_office_tracker_2026,
  title={Ensemble Intelligence for Multi-Asset Portfolio Management},
  author={[Your Name]},
  year={2026},
  url={https://github.com/[your-repo]}
}

📝 License

This project is developed as a semester capstone. All rights reserved.