[Paper] Smart Timing for Mining: A Deep Learning Framework for Bitcoin Hardware ROI Prediction

Source: arXiv - 2512.05402v1

Overview

The paper introduces MineROI‑Net, a Transformer‑based deep‑learning framework that predicts whether buying a new Bitcoin ASIC miner will be profitable, marginal, or unprofitable over the next 12 months. By treating hardware acquisition as a time‑series classification problem, the authors provide the first data‑driven tool to help miners time their purchases amid volatile prices, rapid tech turnover, and Bitcoin’s protocol‑driven revenue cycles.

Key Contributions

• Problem formulation: Cast Bitcoin ASIC acquisition timing into a three‑class time‑series classification task (profitable, marginal, unprofitable).
• Novel architecture: Design of MineROI‑Net, a multi‑scale Transformer that captures short‑ and long‑term patterns in mining profitability indicators.
• Comprehensive dataset: Curated a longitudinal dataset covering 20 ASIC models released from 2015‑2024, spanning multiple market regimes (boom, bust, halving events).
• Strong empirical performance: Achieves 83.7 % overall accuracy and 83.1 % macro F1‑score, with >98 % precision for correctly identifying profitable windows and >93 % precision for spotting unprofitable periods.
• Open‑source release: Full code, pretrained weights, and data preprocessing scripts are publicly available on GitHub, encouraging reproducibility and community extensions.

Methodology

1. Data collection & labeling – Daily market data (BTC price, network difficulty, electricity cost, hash‑rate, block reward) and hardware specs (hash‑rate, power consumption, purchase price) were gathered. For each ASIC model, ROI over the subsequent year was computed; thresholds (ROI ≥ 1, 0 < ROI < 1, ROI ≤ 0) defined the three classes.
2. Feature engineering – Raw time series were transformed into normalized signals (e.g., price‑to‑difficulty ratio, electricity‑adjusted profitability) and fed into the model as a multivariate sequence.
3. Model architecture – MineROI‑Net stacks two Transformer encoder blocks with different temporal receptive fields (short‑term 7‑day windows, long‑term 30‑day windows). Positional encodings preserve order, while a lightweight classification head outputs class probabilities.
4. Training & evaluation – A stratified 5‑fold cross‑validation scheme was used, with class‑balanced loss weighting to mitigate the natural skew toward “marginal” outcomes. Baselines included LSTM, Temporal Convolutional Networks (TCN), and the recent TSLANet.
5. Interpretability – Attention maps were visualized to show which market signals (e.g., difficulty spikes, price drops) drove the model’s decisions, offering actionable insights for miners.

Results & Findings

• The model consistently flags unprofitable periods early, reducing the risk of buying hardware just before a market downturn.
• Attention analysis reveals that spikes in network difficulty combined with falling BTC prices are the strongest predictors of upcoming negative ROI.
• Performance remains robust across different ASIC generations, indicating good generalization to future hardware releases.

Practical Implications

• Decision support for mining farms: Operators can integrate MineROI‑Net into procurement pipelines to schedule purchases during high‑ROI windows, potentially improving cash‑flow stability.
• Risk management for investors: Venture funds and hardware manufacturers can use the model to forecast demand cycles, aligning production and inventory with market profitability.
• Automation potential: The open‑source implementation can be wrapped in a simple API, enabling real‑time alerts (e.g., Slack, email) when a profitable acquisition window opens.
• Energy‑cost optimization: By coupling the model with location‑specific electricity tariffs, miners can decide not only when but also where to deploy new ASICs for maximum ROI.

Limitations & Future Work

• Scope limited to one‑year ROI: Longer‑term profitability (e.g., 2‑3 years) and hardware depreciation are not modeled.
• Assumes static electricity cost: Real‑world miners often negotiate variable rates; incorporating dynamic pricing could improve accuracy.
• Geographic factors omitted: Regulatory changes, tax incentives, and climate‑related cooling costs are outside the current feature set.
• Model interpretability: While attention maps provide some insight, a more rigorous explainability framework (e.g., SHAP for time series) would help users trust the predictions.

Future research could extend the framework to multi‑year horizons, integrate location‑specific cost structures, and explore reinforcement‑learning approaches for end‑to‑end acquisition and deployment strategies.

Authors

• Sithumi Wickramasinghe
• Bikramjit Das
• Dorien Herremans

Paper Information

• arXiv ID: 2512.05402v1
• Categories: cs.LG, cs.AI, cs.CE, cs.NE
• Published: December 5, 2025
• PDF: Download PDF