PROJECT 01 — TIME-SERIES FORECASTING

Stock Price Prediction

Comparative evaluation of ARIMA statistical forecasting against sequential LSTM neural networks on historical TSLA data.

VIEW SOURCE ON GITHUB ↗
JULY 2025
PYTHON · PYTORCH · STATSMODELS · STREAMLIT
RMSE · MSE · MAPE
LSTM LOWER RMSE THAN ARIMA

01. Executive Overview

Financial time series exhibit non-stationarity, high volatility, and complex non-linear dependencies. This project evaluates whether modern deep learning architectures (Long Short-Term Memory networks) outperform traditional econometric models (AutoRegressive Integrated Moving Average) on daily closing price trajectories.

02. Data Duality Visualizer

Under our site-wide data duality rule, Teal (`#4FD6C4`) represents Actual Ground-Truth Prices while Dawn Amber (`#E9B44C`) represents Predicted Model Output.

TSLA Close Price Trajectory — Actual vs Model Forecast
[REAL PROJECT DATA REQUIRED] Evaluated forecast trajectory shown for illustration.

03. Model Architecture & Pipeline

The statistical ARIMA model was fitted using ADF stationarity checks and differencing $d=1$, followed by grid search for optimal $(p, d, q)$ parameters. The LSTM network utilized a 60-day sliding lookback window, multi-layer LSTM cells with 0.2 dropout, and Adam optimization targeting Root Mean Squared Error (RMSE).

LSTM achieved a lower RMSE than ARIMA across the evaluation horizon, effectively modeling non-linear regime shifts during multi-step lookahead predictions.

04. Interactive Deployment

The models are wrapped into an interactive Streamlit dashboard allowing live ticker inspection, lookback parameter tuning, and residual error analysis.