Smart Timing for Mining: A Deep Learning Framework for Bitcoin Hardware ROI prediction

This repository contains the code and dataset accompanying the paper Smart Timing for Mining: A Deep Learning Framework for Bitcoin Hardware ROI Prediction by Sithumi Wickramasinghe (PhD candidate), Prof. Bikramjith Das, and Prof. Dorien Herremans.

We propose MineROI-Net a Transformer-based architecture for timing Bitcoin mining hardware purchases.
Given a 30- or 60-day window of machine, market, and network features, the model predicts whether buying a specific ASIC miner on a given day will be:

• Unprofitable (ROI ≤ 0)
• Marginal (0 < ROI < 1)
• Profitable (ROI ≥ 1 within 365 days)

This repository contains the PyTorch implementation, data preprocessing pipeline, and experiments used in our paper.

🔥 Live demo online!

---

Overview

Bitcoin mining is capital-intensive and highly sensitive to market cycles, halving events, hardware efficiency, and electricity prices. MineROI-Net formulates the hardware acquisition timing problem as a multi-class time series classification task.

The model combines:

1. Spectral Feature Extractor

   • FFT-based layer with learnable complex weights to highlight important frequency components (halving cycles, difficulty adjustment cycles, etc.).

2. Channel Mixing Module

   • Squeeze-and-Excitation style feature re-weighting that captures cross-feature interactions (e.g., price vs. electricity cost dominance) in a lightweight way.

3. Transformer Encoder

   • Standard multi-head self-attention over the processed sequence, followed by global pooling and a classification head.

MineROI-Net outperforms LSTM-based and TSLANet baselines on data from 20 ASIC miners (2015–2024), achieving strong accuracy and macro-F1 while being economically well-behaved (very high precision for profitable and unprofitable periods).

---

Data Collection Pipeline

The data_collection/ folder contains 6 sequential scripts that download, process, and prepare datasets for 20 ASIC mining machines across three regions (Texas, China, Ethiopia).

Note: Due to data-source restrictions, we do not redistribute raw ASIC pricing data. Follow the data sources in Section 4.1 of the paper and these scripts to reconstruct the dataset.

---

Pipeline Steps

Step 1: Download Blockchain Data

Script: 1_download_blockchain_data.py

Downloads Bitcoin price, network difficulty, transaction fees, hashrate, and miners revenue from blockchain.info API.

Date range: Jan 17, 2009 - Sep 23, 2025
Output: blockchain_data.csv

---

Step 2: Download ASIC Price Data

Script: 2_download_asic_price_data.py

Downloads historical prices for 20 ASIC miners from Hashrate Index API.

Requires: API key from hashrateindex.com/api
Date range: Jan 22, 2018 - Sep 21, 2025
Output: asic_prices.csv, miner_data/ (individual CSVs)

---

Step 3: Prepare Electricity Data

Script: 3_electricity_data.py

Prepare electricity price data for three regions. See Section 4.1 in paper for data sources.

Required files:

• texas_residential_daily_df.csv
• china_electricity_prices_daily.csv
• ethiopia_electricity_prices_daily.csv

Format: date (YYYY-MM-DD) and price (USD/kWh) columns

---

Step 4: Prepare Miner Datasets

Script: 4_prepare_miner_dataset.py

Combines blockchain data with ASIC specifications for all 20 miners.

Calculations:

• Machine specifications (hashrate, power, efficiency, release date)
• Block rewards (Bitcoin halving schedule)
• Machine age and days since halving
• Daily revenue potential

Output: full_feature_data.csv

---

Step 5: Calculate ROI by Country

Script: 5_roi_country.py

Calculates 12-month forward ROI for each miner in each region.

Output: roi_texas.csv, roi_china.csv, roi_ethiopia.csv

---

Step 6: Create Target Variable

Script: 6_create_target.py

Engineers time-series features and creates ROI classification target.

ROI Categories:

• 0: Unprofitable (ROI < 0)
• 1: Marginal (0 ≤ ROI < 1)
• 2: Profitable (ROI ≥ 1)

Output: final_texas.csv, final_china.csv, final_ethiopia.csv

---

Quick Start

Run all steps in sequence:

python 1_download_blockchain_data.py
python 2_download_asic_price_data.py --api-key YOUR_API_KEY
python 3_electricity_data.py
python 4_prepare_miner_dataset.py
python 5_roi_country.py
python 6_create_target.py

---

Final Dataset

Each final_*.csv contains:

Key Features:

• date, bitcoin_price, difficulty, fees, hashrate, revenue
• machine_price, machine_hashrate, power, efficiency
• block_reward, age_days, days_since_halving
• Revenue_Potential, electricity_rate, machine_name

Target:

• roi_category_id: 0 (Unprofitable), 1 (Marginal), 2 (Profitable)

---

Dataset Sample

date	bitcoin_price	difficulty	fees	hashrate	revenue	machine_price	machine_hashrate	power	efficiency	block_reward	age_days	days_since_halving	Revenue_Potential	machine_name	electricity_rate	roi_category_id
2024-09-18	60304.22	9.27E13	7.02	6.36E8	2.64E7	833.94	226	6554	29	3.125	1265	151	9.245	m53	0.0767	0
2024-09-19	61683.91	9.27E13	9.94	5.62E8	2.50E7	833.94	226	6554	29	3.125	1266	152	9.457	m53	0.0767	0
2024-09-20	62938.20	9.27E13	9.51	6.68E8	2.93E7	833.94	226	6554	29	3.125	1267	153	9.649	m53	0.0767	0
2024-09-21	63213.19	9.27E13	6.08	6.22E8	2.70E7	833.94	226	6554	29	3.125	1268	154	9.691	m53	0.0767	0

---

MineROI-Net Model Training

The models folder consists of two main components:

1. dataloader.py - Prepares time-series data for model training
2. transformer_final_split.py - Trains and evaluates the transformer model

---

Script: dataloader.py

Prepares the final datasets for training by creating time-series windows and splitting data.

from dataloader import run_all_preprocessing

# Create time-series windows and prepare datasets
run_all_preprocessing(SEQ_LEN=30)

What it does:

• Creates 30-day sliding windows from time-series data
• Splits data: 80% train, 20% val/test (time-based)
• Scales features with MinMaxScaler
• Combines data from all three regions
• Converts to transformer format [Batch, Length, Channels]

Output: seq_30/train_trans.pt, val_trans.pt, test_trans.pt

---

Script: transformer_final_split.py

python transformer_final_split.py

What it does:

• Trains transformer classifier on preprocessed data
• Predicts ROI categories: 0 (Loss), 1 (Partial), 2 (Profitable)
• Uses early stopping and saves best checkpoint
• Reports accuracy, precision, recall, F1-score

Output: checkpoints/best_model.pth

---

📄 Citation

If you find this work useful, please cite our paper:

@article{wickramasinghe2025smart,
  title   = {Smart Timing for Mining: A Deep Learning Framework for Bitcoin Hardware ROI Prediction},
  author  = {Wickramasinghe, Sithumi and Das, Bikramjit and Herremans, Dorien},
  journal = {arXiv preprint arXiv:2512.05402},
  year    = {2025},
  url     = {https://arxiv.org/abs/2512.05402}
}