📊 Business KPI Prediction from Financial Data (PyTorch)

---

1. Introduction

This repository implements a multi-task deep learning model for predicting key business performance indicators (KPIs) from tabular financial/accounting data.
The model jointly learns three predictive tasks:

• Revenue Growth — regression
• Risk Score — regression
• Customer Churn — binary classification

By training on all three tasks simultaneously, the network captures richer patterns and relationships in financial data.

---

2. Motivation

Conventional financial forecasting methods (linear models, rule-based systems, spreadsheets) struggle with nonlinear dependencies and noisy real-world data.

This project demonstrates how multi-task learning (MTL) can serve as a scalable, modern approach to forecasting KPIs, providing:

• One model → multiple outputs
• Feature integration across categorical, numerical, and temporal inputs
• Extensible baseline for real-world datasets

---

3. Dataset

• Source: synthetic_financial_data_bukharii.csv (synthetic but realistic)
• Records: ~34,000 rows (500 companies × 68 months)
• Features:
  • Numerical: revenue, gross_profit, operating_margin, debt_ratio, log_revenue
  • Categorical: industry, region, company_size
  • Temporal: customer_tenure, date
• Targets: revenue growth, risk score, churn indicator

⚠️ Note: Dataset is synthetic and designed for experimentation.

---

4. Model Architecture

• Base network: Fully connected layers with BatchNorm + Dropout
• Output heads:
  • Revenue growth → Linear regression head (MSELoss)
  • Risk score → Linear regression head (MSELoss)
  • Churn → Binary classification head (BCEWithLogitsLoss)
• Framework: PyTorch

---

5. Preprocessing Pipeline

All preprocessing steps are modularized in preprocessing.py:

• Missing value handling
• Feature scaling (with scaler.pkl)
• Label/categorical encoding (with encoders.pkl)
• Automatic column validation during inference

This ensures consistency between training and deployment.

---

6. Training Setup

• Split: 2020–2024 → training | 2024–2025 → validation/testing
• Optimizer: Adam (lr = 5e-5)
• Loss: MSE(revenue) + MSE(risk) + BCE(churn)
• Batch size: 128
• Epochs: 100

Sample Training Output

Epoch	Train Loss	Validation Loss
1	1.1861	1.2510
5	0.7073	0.7825
10	0.6764	0.6891
15	0.5981	0.6107
20	0.6054	0.6232

---

7. Results

• Model shows stable convergence on all three tasks.
• Training and validation losses are closely aligned → low overfitting on synthetic data.
• Demonstrates feasibility of joint KPI forecasting using deep learning.

(Future improvement: report R² for regression tasks and AUC/F1 for churn classification.)

---

8. Deployment (Streamlit App)

This repository includes a Streamlit web app for interactive predictions.

Run Locally

Make sure you have trained the model and saved artifacts (finance_model.pth, scaler.pkl, encoders.pkl) in the models/ folder.

# Step 1: Train the model
python train.py

# Step 2: Validate model performance
python validate.py

# Step 3: Launch Streamlit app
streamlit run server/app.py