NVIDIA Stock Forecasting and Analysis

This repository provides an end‑to‑end pipeline—data cleaning, feature engineering, classical econometrics, sentiment fusion, and deep learning—for NVIDIA (NVDA) price forecasting.

Table of Contents

1. Project Overview
2. Data
3. Environment and Setup
4. Usage
5. Methods
   • Data Pre-processing & Feature Selection
   • Classical Time-Series Models
   • Sentiment Integration
   • Hybrid RF → LSTM-Attention
6. Results
7. Figures
8. References
9. License

---

Project Overview

1. Forecast NVDA prices with ARIMA, SARIMAX, GARCH, and a hybrid Random‑Forest ➔ Bidirectional LSTM + Attention network.
2. Explain drivers such as S&P 500, Google, Microsoft, Intel, macro factors, and a news‑based ImpactScore.
3. Compare each layer of model complexity against a simple T‑2 baseline (price from two days prior).

---

Data

Below is a list of each data file contained in the Data directory:

• Backup.csv
• Database.csv.csv
• cleaned_data.csv
• merged_data.csv
• merged_data2.csv
• merged_with_impact.csv
• merged_with_impact_score.csv
• nvidia_events_filtered.csv

Each file is used at different stages of the data processing and analysis pipeline. Please refer to the specific sections of the project documentation for details on how each file is utilized

---

Environment and Setup

conda create -n nvda_env python=3.9
conda activate nvda_env
pip install -r requirements.txt

Key libraries – pandas • numpy • scikit-learn • statsmodels • arch • tensorflow (keras)

---

Usage

1 . Clone

git clone https://github.com/YourUsername/NVIDIA-Forecasting.git
cd NVIDIA-Forecasting

2 . Drop data into data/

mkdir -p data
# place cleaned_data.csv and merged_with_impact_score.csv here

3 . Run full notebook‑style script (quick start)

python "Total Code.py"

4 . Modular workflow (recommended)

Step	Script	Core Techniques
① Pre‑processing	src/data_prep.py	drop NA/∞, StandardScaler, add 20‑/80‑day MA
② Feature ranking	src/feature_select.py	Pearson corr, PCA 95 %, LassoCV, Random Forest
③ ARIMA	src/arima.py	ADF test → ARIMA(4,1,0) rolling forecast
④ SARIMAX & GARCH	src/sarimax_garch.py	exog = [SP500_log, ImpactScore, INTC_ret]
⑤ Sentiment scrape	src/news_sentiment.py	VADER / FinBERT → daily ImpactScore
⑥ Deep model	src/lstm_attention.py	RF meta‑feature ➔ 2×50 bi‑LSTM + Attention
⑦ Plots	src/visualization.py	saves all figures to docs/img/

Data Pre‑processing & Feature Selection

We start by removing NaN/Inf rows and standardising all numerical columns.
Two complementary feature‑ranking tracks are applied:

Technique	Purpose	Outcome
PCA (95 % var)	Orthogonalise & compress	9 principal components retained
LassoCV + Random Forest	Sparse, non‑parametric importance	Top drivers: SP500, MSFT_Adj_Close, ImpactScore, 20‑/80‑day MA

These features feed every downstream model to ensure consistency and avoid look‑ahead bias.

---

Classical Time‑Series Models

ARIMA

The ADF test rejects the unit‑root hypothesis after one differencing, leading to an (4,1,0) specification chosen via AIC grid search.
The autocorrelation structure is visualised below:

Both plots confirm strong short‑memory up to three lags, justifying the AR term.

SARIMAX

Seasonality (12‑month) and exogenous regressors—SP500_log, ImpactScore, INTC_ret—are introduced in a SARIMAX(1,0,1)(0,1,1,12) framework.

Left panel: fitted vs. observed shows tight tracking.
Right panel: standardised residuals & Q‑Q plot indicate near‑normality with mild tail risk—later captured by GARCH.

GARCH

A GARCH(1,1) layer is fitted to log‑return residuals, reducing volatility clustering and yielding a log‑likelihood of −2156 (↑ vs. ARIMA).

---

Sentiment Integration

Daily news headlines are scored by VADER and FinBERT; scores are averaged into an ImpactScore that enters SARIMAX and LSTM as a leading indicator.

Scatter illustrates a Pearson‑r = 0.43 between ImpactScore and next‑day return.
Event overlay shows price jumps aligning with major positive (green) and negative (red) news.

---

Hybrid RF → LSTM‑Attention

1. Random Forest predicts next‑day price to create a meta‑feature.
2. Neural net architecture:

Input (20 features, T‑2 window)
     └─ bi‑LSTM (50) ─┐
     └─ bi‑LSTM (50) ─┘ → Attention → LSTM (50) → Dense(1)

3. Regularisation: Dropout 0.3, L2 0.01, EarlyStopping (patience 10).

The plot compares Actual (blue), LSTM (red), and the naive T‑2 baseline (green).

---

Results

Model	Inputs	Best Test Metric	T‑2 Baseline
ARIMA(4,1,0)	Close	MSE 34.2	—
SARIMAX	Close + exog	MSE 22.5	—
GARCH(1,1)	log‑σ²	LLH ↑ −2156	—
LSTM‑Attention	20 features, time_steps = 2	MAE 2.42	3.44

• 32 % MAE reduction from baseline to LSTM.
• SARIMAX halves ARIMA’s error by injecting macro + sentiment.
• Residual heavy tails in SARIMAX are mostly neutralised after the GARCH layer.

References

• Xiao, Q., & Ihnaini, B. (2023). Stock trend prediction using sentiment analysis. PeerJ Computer Science.
• Yahoo Finance – NVDA
• Mnih, V. et al. (2016). Asynchronous Methods for Deep Reinforcement Learning. arXiv:1607.01958.

---

Thank you for visiting this project. If you have any questions or suggestions, feel free to open an issue or submit a pull request.