PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Sithumi Wickramasinghe, Bikramjit Das, Dorien Herremans

arXiv:2512.05402v2cs.LGcs.AIcs.CEcs.NE

TL;DR

Mining operators lack formal guidance on when to acquire ASIC hardware amid volatile costs and cyclical profitability. The paper classifies purchase outcomes over a one-year ROI horizon, and MINEROI-NET outperforms the evaluated baselines on the reported results.

  • Problem

    Formal guidance on hardware-acquisition timing is lacking despite cyclical profitability and volatile costs, while poorly timed purchases can sharply lengthen payback.

    Source: §1 · p. 2

  • Method

    The paper uses a data-driven time-series classification framework to predict profitable, moderate, or negative ASIC-purchase ROI over a 365-day horizon.

    Source: §1 · p. 2

  • Results

    MINEROI-NET outperforms all four baselines, reaching 83.2% accuracy and a macro F1-score of 83.5% with the best-performing 60-day look-back configuration.

    Source: §6 · p. 15

  • Takeaways & Limitations

    The authors conclude that MINEROI-NET offers practical value for capital-intensive mining decisions and informed hardware-investment timing.

    Source: §6 · p. 16

    The setup assumes periodic BTC-to-fiat conversion over a fixed 365-day horizon; alternative selling strategies and other horizons remain future-work directions.

    Source: §3.1 · p. 4, §6 · p. 15

Abstract

from arXiv · show

Bitcoin mining hardware acquisition requires strategic timing due to volatile markets, rapid technological obsolescence, and protocol-driven revenue cycles. Despite mining's evolution into a capital-intensive industry, there is little guidance on when to purchase new Application-Specific Integrated Circuit (ASIC) hardware, and no prior computational frameworks address this decision problem. We address this gap by formulating hardware acquisition as a time series classification task, predicting whether purchasing ASIC machines yields profitable (Return on Investment (ROI) >= 1), marginal (0 < ROI < 1), or unprofitable (ROI <= 0) returns within one year. We propose MineROI-Net, an open-source Transformer-based architecture designed to capture multi-scale temporal patterns in mining profitability. Evaluated on data from 20 ASIC miners released between 2015 and 2024 across diverse market regimes, MineROI-Net outperforms recurrent, convolutional, and attention-based baselines, achieving 83.2% accuracy and 83.5% macro F1-score. The model demonstrates strong economic relevance, achieving 97.8% precision in detecting unprofitable periods and 81.5% precision in detecting profitable ones, while avoiding misclassifying profitable scenarios as unprofitable and vice versa. These results indicate that MineROI-Net offers a practical, data-driven tool for timing mining hardware acquisitions, potentially reducing financial risk in capital-intensive mining operations.

1. Introduction

The introduction identifies hardware-acquisition timing as an underexplored decision for miners facing cyclical profitability and costly, rapidly depreciating equipment. It presents a data-driven classification framework and MINEROI-NET as a response to that gap.

1. Introduction

  • No formal framework guides when miners should expand capacity despite hardware obsolescence, electricity-price volatility, and protocol-driven revenue cycles.Source: §1 · p. 2
  • Peak-market Antminer S19j Pro buyers saw expected payback rise from approximately 13 months to 107 months after mining margins collapsed.Source: §1 · p. 2
  • The framework predicts profitable, moderate, or negative ROI from a hardware purchase using machine, Bitcoin-market, price, and network data over a 365-day horizon.Source: §1 · p. 2
  • MINEROI-NET reports 83.2% accuracy and 97.8%/81.5% precision for unprofitable/profitable predictions, benchmarked against LSTM-based, transformer-based, and TSLANet architectures.Source: §1 · p. 2

2. Related work

Mining economics research describes cost, revenue, and protocol drivers, while deep-learning studies chiefly address cryptocurrency market prediction. The paper positions acquisition timing as a distinct investment decision with an explicit ROI horizon.

2.1. Mining economics and profitability drivers

  • Electricity accounts for more than 80% of miners’ cash-based operating expenses, while hardware renewal, rewards, halvings, fees, and network conditions also shape profitability.Source: §2.1 · pp. 2–3
  • Prayoga et al. (2025) predict daily device income, but do not include acquisition costs, define a fixed ROI horizon, or test generalization across temporal shifts or new hardware.Source: §2.1 · p. 3

2.2. Deep learning for cryptocurrency markets and time-series decision support

  • Cryptocurrency deep-learning work spans recurrent LSTMs and multimodal inputs, while Transformers model long-range dependencies and convolutional networks capture local temporal patterns.Source: §2.2 · pp. 3–4
  • MINEROI-NET frames hardware acquisition timing as a classification task aligned with capital commitment and a fixed one-year ROI horizon, unlike prior market-prediction tasks.Source: §2.2 · p. 4

3. Approach and model

The approach defines purchase-date sequences and a one-year ROI target, computes returns through cash-flow accounting, and bins outcomes into three classes. MineROI-Net combines spectral modulation, channel weighting, and a Transformer encoder for classification.

3. Approach and model

  • Each sample pairs a machine and purchase date with a past L-day window of machine, market, and network features to predict its one-year ROI class.The task treats acquisition timing as multi-class time-series classification.Source: §3 · p. 4

3.1. ROI computation and class labeling

  • ROI is modeled over 365 days from mining cash flows relative to initial hardware cost, accounting for acquisition, electricity, and maintenance.The baseline assumes daily conversion of mined Bitcoin to fiat; electricity is fixed within each scenario and tested from 0.01–0.20 USD/kWh, while annual maintenance is set to 5% of purchase price.Source: §3.1 · pp. 4–5
  • One-year ROI is labeled unprofitable at ROI ≤ 0, marginal at 0 < ROI < 1, and profitable at ROI ≥ 1.These classes correspond to capital loss, partial recovery, and full capital recovery with at least a 100% return, respectively.Source: §3.1 · p. 6

3.2. Proposed model architecture

  • MineROI-Net processes mining-profitability sequences through a spectral feature extractor, channel-mixing module, and Transformer encoder.The architecture is designed to account for temporal dependencies, regime shifts, and patterns spanning daily fluctuations to multi-year cycles.Source: §3.2 · p. 6

3.2.1. Frequency-Specific Spectral Modulation

  • The spectral block transforms each feature’s time series into the frequency domain, learns frequency-specific modulation, and maps the modulated spectrum back to time.This lets the model amplify or attenuate periodic components, complementing attention-based temporal modeling of mining’s multi-timescale cycles.Source: §3.2.1 · pp. 6–7

3.2.2. Channel Mixing Module

  • Channel mixing pools each feature over time and uses a shared bottleneck to produce context-dependent importance scores that reweight features.High-scored features are amplified and less relevant features suppressed, addressing cross-feature relationships that vary across market regimes.Source: §3.2.2 · p. 7

3.2.3. Transformer Encoder

  • The Transformer encoder receives projected inputs with sinusoidal positional encodings and processes them through stacked layers with self-attention, feed-forward networks, normalization, and residual connections.Its final temporal representation is formed from the encoded sequence.Source: §3.2.3 · p. 8

3.2.4. Classification and Training

  • Global average pooling converts the final encoded sequence to a fixed-length vector, and a two-layer head predicts probabilities for the three ROI classes.Training uses weighted cross-entropy with inverse-frequency class weights and label smoothing with ε = 0.1.Source: §3.2.4 · p. 8

4. Experimental setup

The experimental setup describes the dataset and input construction, baseline models, and time-ordered evaluation and hyperparameter-selection procedures.

4.1. Dataset and input construction

  • The daily dataset spans October 2015–September 2025, with ROI labels through September 2024, and combines 20 ASICs, market and network variables, and electricity scenarios of 0.01–0.20 USD/kWh.Machine attributes include specifications and purchase prices; the evaluation covers multiple bitcoin market cycles, difficulty regimes, and hardware generations.Source: §4.1 · pp. 8–9

4.2. Baseline models

  • MineROI-Net combines FFT-based spectral features and channel mixing with a Transformer; comparisons include LSTM, Transformer, and TSLANet baselines.The LSTM-with-blocks baseline holds spectral extraction and channel mixing constant while replacing the Transformer encoder with recurrent layers.Source: §4.2 · p. 9

4.3. Time-based data split and expanding-window cross-validation

  • An 80/20 chronological split places the training boundary at May 2023; S21 appears only in the test period, assessing generalization to unseen hardware and future market regimes.Samples are machine-isolated and temporally ordered to prevent future information entering past windows.Source: §4.3 · pp. 9–10, Figure 2 · p. 10
  • Three expanding-window validation phases represent bear, range-bound, and bull markets; models are selected by validation macro F1 and then evaluated on the untouched June 2023–September 2024 test window.The final model configurations selected through rolling cross-validation are listed in Table 1.Source: §4.3 · p. 10, Figure 3 · p. 11, Table 1 · p. 11

5. Results

MINEROI-NET leads both test metrics across the 30- and 60-day settings, with its strongest reported results at 60 days. Longer context improves balanced class performance, and the 60-day model makes no direct unprofitable-to-profitable or profitable-to-unprofitable confusions.

5.1. Final test performance with 30-day look-back

  • Attention-based models outperform purely recurrent ones, while frequency-aware and cross-feature modules improve LSTM performance over Vanilla LSTM.Source: §5.1 · p. 11
  • At 60 days, MINEROI-NET achieves 0.832 accuracy and 0.835 macro-F1, the best performance across both look-back settings.It also leads both metrics in the 30-day setting.Source: Table 2 · p. 12, §5.2 · p. 12, §5.1 · p. 12

5.2. Influence of the look-back window

  • Longer context improved every model family: Vanilla LSTM accuracy rose 0.611→0.670, LSTM with blocks 0.637→0.717, Vanilla Transformer 0.770→0.798, and TSLANet 0.714→0.784.Source: §5.2 · p. 12

5.3. Error analysis and class-wise performance

  • At 60 days, unprofitable recall reaches 0.868 from 0.756 and no U↔P errors occur; marginal cases remain the main boundary uncertainty.The reported confusion matrices and per-class metrics cover both look-back settings.Source: §5.3 · pp. 12–13, Table 3 · p. 13

5.4. Ablation study of spectral modulation and channel mixing

  • On the 60-day test, the full model reaches 0.859 macro-F1, beating vanilla by +0.090 and single-module variants by +0.073 and +0.055.Channel mixing alone adds +0.035 macro-F1, versus +0.017 for spectral modulation alone.Source: Table 4 · p. 14, §5.4 · p. 14

5.5. Sensitivity to alternative conversion-frequency assumptions

  • Under monthly relabeling, accuracy and macro-F1 shift from 0.858 and 0.859 to 0.854 and 0.856; most labels persist, with changes only between adjacent classes.96.3% of unprofitable, 95.3% of marginal, and 99.6% of profitable cases retain their labels.Source: §5.5 · pp. 14–15, Table 6 · p. 15

6. Conclusion

The conclusion presents MINEROI-NET as an open-source model for predicting one-year Bitcoin mining hardware ROI classes. It reports strong benchmark results and outlines future directions for extending the framework.

  • MINEROI-NET predicts one-year hardware-acquisition ROI as profitable (ROI ≥1), marginal (0 < ROI < 1), or unprofitable (ROI ≤0).Source: §6 · p. 15
  • 83.2% accuracy and 83.5% macro F1-score; MINEROI-NET consistently outperforms Vanilla LSTM, LSTM with blocks, Vanilla Transformer, and TSLANet.Source: §6 · p. 15
  • 97.8% precision for unprofitable periods and 81.5% for profitable ones, without misclassifying profitable scenarios as unprofitable or vice versa.Source: §6 · p. 15
  • The test set includes S21, absent from training; spectral modulation and channel mixing together contribute +0.090 macro F1 over the vanilla Transformer backbone.Source: §6 · p. 15
  • Performance remains stable under monthly Bitcoin conversion assumptions, with accuracy and macro F1-score decreasing only marginally to 85.4% and 85.6%, respectively.Source: §6 · p. 15
  • Future work includes weekly or holding-based conversion, 6- or 18-month ROI horizons, other cryptocurrency mining ecosystems, and regression formulations.Source: §6 · pp. 15–16

Related papers

Loading 2512.05402v2…