redr1g/final-thesis-experiments

Time-Series Forecasting for Virtual Item Prices

This repository contains trained machine learning and deep learning models for forecasting the price direction of virtual assets based on time-series data.

Best Model

The best performance was achieved by the LSTM neural network on the Kneedle (MAE <= 0.416) data pool, yielding a metric of ROC-AUC = 0.6567.

Final model architecture:

• LSTM(128) – recurrent layer, tanh activation
• Dropout(0.3)
• Dense(64, ReLU)
• Dense(1, sigmoid) – probability of price increase

Repository Structure

• lstm_final_kneedle.keras — the main optimized model.
• lstm_final_kneedle_cosmetics.keras — model for the alternative dataset (appendix).
• best_experiments/ — the best models across 14 different algorithms (CatBoost, XGBoost, TabNet, etc.) after pool optimization.

Benchmarking

Model	Best Pool	ROC-AUC
LSTM	Kneedle (MAE ≤ 0.416)	0.6567
CatBoost	RF Gating (t = 0.7)	0.6446
XGBoost	ENN (t = 0.7)	0.6405
TabNet	RF Gating (t = 0.7)	0.6398
LightGBM	DROP3 (t = 0.7)	0.6339
Decision Tree	RF Gating (t = 0.7)	0.6296
Random Forest	RF Gating (t = 0.7)	0.6294
MLP	ENN (t = 0.7)	0.6292
Logistic Regression	DT Filter (t = 0.5)	0.6235
SVM	ENN (t = 0.7)	0.6178
KNN	Kneedle (MAE ≤ 0.416)	0.5844
Naive Bayes	DT Filter (t = 0.5)	0.5731
1D-CNN	ENN (t = 0.7)	0.5367
AutoARIMA	DT Filter (t = 0.7)	0.5027

Inference

from huggingface_hub import hf_hub_download
from tensorflow.keras.models import load_model

# Load the best model
model_path = hf_hub_download(
    repo_id="redr1g/final-thesis-experiments", 
    filename="lstm_final_kneedle.keras"
)
model = load_model(model_path)