FinCast Pro is a professional forecasting platform built in Python and Streamlit. It runs four independent models — Prophet, ARIMA/SARIMA, XGBoost, and Monte Carlo simulation — on any time series, then blends them into an ensemble forecast. The app includes walk-forward backtesting, residual diagnostics, bias–variance decomposition, regime detection, and full Excel/CSV export. It is designed for analysts, data scientists, and portfolio managers who need production-quality forecasts without writing code.

Upload any CSV — flexible column detection auto-maps date/time/ds and value/close/price/revenue/y columns; no specific naming required

Yahoo Finance integration — enter any ticker (AAPL, BTC-USD, MSFT, ETH-USD, ^GSPC) and pull up to 10 years of live OHLCV data

Built-in sample datasets — 48-month monthly revenue with seasonality and injected anomaly, plus a 500-day GBM stock simulation

Auto frequency detection — infers Monthly / Quarterly / Weekly / Daily (Business) from the median gap between dates

IQR anomaly detection — flags outliers with a configurable IQR multiplier (1×–4×), overlaid on the historical chart

Winsorisation toggle — clips anomalies to IQR bounds before fitting so models train on clean data; raw data is preserved in all charts and exports

Regime detection — CUSUM structural break detection identifies regime shifts with amber overlays on the historical series

Model Algorithm Confidence Interval Prophet Meta's additive decomposition with Fourier seasonality and trend changepoints 90% posterior uncertainty ARIMA / SARIMA Statsmodels SARIMAX(1,1,1) with automatic seasonal order detection 90% analytical CI XGBoost Gradient-boosted trees on a 12-lag feature window with recursive multi-step forecasting Expanding uncertainty bands Monte Carlo Geometric Brownian Motion with scenario-adjusted drift and volatility P10–P90 percentile fan Ensemble Equal-weight blend of all active model point forecasts Averaged CI bands

Five scenario presets: Stress · Bear · Base · Bull · Upside adjusting GBM drift (μ) and volatility (σ) multipliers

VaR (95%) and CVaR / Expected Shortfall computed from the Monte Carlo terminal distribution

Full percentile table P1–P99 with % change vs. last observed value

Probability of loss and probability of gain > 10%

Monte Carlo fan chart with 100 individual simulated paths overlaid

Walk-forward backtesting — 5-fold rolling-origin out-of-sample validation; each fold trains on an expanding window and tests on the next held-out period

Residual diagnostics — time-series residual plot and ACF bar chart per model with 95% significance bands

Bias² vs. Variance decomposition — bar chart separating systematic error (underfitting) from variance (overfitting)

Model leaderboard — ranked by in-sample MAPE with 🥇🥈🥉 medals and ✦ Best Fit annotation

In-sample fit overlay — all fitted models overlaid against actual data for visual inspection

@st.cache_data on all four model runners — second run with identical inputs returns instantly from cache

XGBoost uses tree_method="hist" — fastest available tree builder

Monte Carlo is fully vectorised with NumPy — no Python loop over simulations

Prophet configured with uncertainty_samples=200 and n_changepoints=15 — ~5× faster than defaults

ARIMA uses L-BFGS optimiser with maxiter=75 — 3–4× faster than Newton default

File Contents Forecast CSV All models · lower / upper CI · ensemble Historical CSV Raw series with anomaly flag column Excel Report (multi-sheet) Historical · Forecast · Model Accuracy · Walk-Forward · MC Summary

Python 3.10 or later

pip

The app opens at http://localhost:8501.

Fork this repository to your GitHub account

Go to share.streamlit.io → New app

Select your fork → branch main → main file app.py

Click Deploy

Streamlit Cloud reads requirements.txt and .streamlit/config.toml automatically. No secrets or environment variables are required for any feature.

Your file needs at minimum two columns — a date column and a numeric value column. Column names are detected automatically:

Accepted date column names Accepted value column names ds, date, time, timestamp, period y, value, close, price, revenue, sales

If no column matches either list, the app uses the first two columns and displays the mapping it applied. Minimum recommended series length: 24 observations for monthly data, 8 for quarterly, 104 for weekly, 500 for daily.

Select Yahoo Finance as the data source in the sidebar

Enter a ticker symbol — e.g. AAPL, BTC-USD, GLD, ^GSPC, EURUSD=X, NVDA

Choose a lookback period: 1y · 2y · 5y · 10y

Frequency is auto-detected as Daily (Business); override in the Frequency selector if needed

Choose your Data Source and load data

Confirm or override the Frequency (auto-detected for uploads and Yahoo data)

Set the Forecast Horizon — number of periods ahead

Toggle individual Models on/off

Optionally enable Winsorise outliers to clean the training data

Choose a Scenario for Monte Carlo

Toggle Show confidence intervals and Show regime breaks as desired

Click ▶ RUN FORECAST ENGINE

Each model shows a live status card (Queued → Running → Done with MAPE). Results appear as soon as all models complete. Clicking Run again with unchanged inputs is instant due to caching.

In-sample MAPE — how well models fit the training data. Lower is better but can be misleading if the model is overfitting.

Walk-forward OOS MAPE — how well models forecast unseen data across 5 rolling folds. This is the number to trust for deployment decisions.

ACF plot — bars outside the dashed CI lines indicate autocorrelated residuals, meaning the model has not captured all signal.

Bias² vs Variance — dominant red (bias²) means the model is systematically under/over-forecasting; dominant blue (variance) means it is overfitting.

Model Parameter Value Rationale Prophet yearly_seasonality 6 Fourier terms Faster; sufficient for most business series Prophet n_changepoints 15 Reduced from default 25; reduces overfitting Prophet uncertainty_samples 200 Reduced from default 1 000; 5× speed improvement Prophet changepoint_prior_scale 0.05 Conservative; prevents trend drift ARIMA Order (1,1,1) Robust default for integrated financial series ARIMA Seasonal order Auto Enabled when len(s) ≥ 3 × seasonal_period ARIMA Optimiser L-BFGS 3–4× faster than Newton ARIMA maxiter 75 Sufficient with L-BFGS XGBoost n_estimators 150 Diminishing returns beyond this for typical TS lengths XGBoost tree_method hist Fastest available; equivalent accuracy XGBoost max_depth 3 Shallower reduces overfitting on small datasets XGBoost Lag window 12 12 lagged values + rolling mean/std/min/max as features Monte Carlo Paths 500–5 000 User-configurable; vectorised via NumPy Monte Carlo Model GBM Log-normal returns, seed 99 for reproducibility

Scenario Drift multiplier (μ×) Volatility multiplier (σ×) Stress 0.3 1.5 Bear 0.3 1.5 Base 1.0 1.0 Bull 1.5 0.65 Upside 1.5 0.65

Metric Formula CAGR (last / first) ^ (1 / years) − 1 × 100 Max Drawdown min((s − cummax(s)) / cummax(s)) × 100 Sharpe Ratio (annualised_return − 0.05) / annualised_std VaR 95% 5th percentile of Monte Carlo terminal distribution CVaR / ES Mean of simulated terminal values ≤ VaR MAPE `mean( sMAPE `mean(2 RMSE sqrt(mean((actual − forecast)²))

Each split trains on df[:train_end] and forecasts step periods. Results are cached per (data_hash, freq) so repeated visits to the Validation tab do not re-run.

Package Version Purpose streamlit ≥ 1.30 Web application framework plotly ≥ 5.0 Interactive charts pandas ≥ 2.0 Data manipulation numpy ≥ 1.24 Numerical computation prophet ≥ 1.1 Additive time series model (Meta) statsmodels ≥ 0.14 ARIMA / SARIMA xgboost ≥ 2.0 Gradient boosted trees yfinance ≥ 0.2 Yahoo Finance data fetching openpyxl ≥ 3.1 Excel export engine scikit-learn ≥ 1.3 Utility functions

Multi-series support — upload a CSV with multiple value columns and forecast each independently

LSTM / Temporal Fusion Transformer — deep learning baseline for comparison

Auto-ARIMA — AIC/BIC grid search over (p, d, q) order combinations

API key support via st.secrets — user-provided Alpha Vantage / FRED / Quandl keys for alternative data

PDF report export — one-click professional report with all charts embedded

Mobile layout optimisation — responsive sidebar collapse for small screens

Forecast accuracy tracking — save past runs and compare forecast vs. actuality over time

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

Commit convention: feat: · fix: · docs: · refactor: · perf:

This project is licensed under the MIT License. See LICENSE for full terms.

GitHub: @Imaad18

Repository: github.com/Imaad18/fincast-pro