Dynamic cointegration screening identifies which crypto pairs remain mean-reverting, and an ensemble of deep networks plus LSTM produces the most accurate spread forecasts and timely backtested buy/sell signals; however, robustness to trading frictions and regime shifts is not established.

Main Finding

A dynamic pairs-trading framework that combines dynamic Johansen co-integration with deep learning (DNN and LSTM) and a Dynamic Weighted Ensemble (DWE) yields superior real-time spread forecasts and trading signals for cryptocurrency pairs. The DWE outperforms static ensemble methods, LSTM effectively captures temporal (proportional) dynamics, and only pairs exhibiting time-varying (dynamically coherent) cointegration are suitable for mean-reversion strategies.

Key Points

• Scope and data: Six cryptocurrency pairs examined (BTC–ETH, BTC–LTC, BTC–XRP, ETH–LTC, ETH–XRP, LTC–XRP) using daily prices from 2018-01-02 to 2025-10-31 (N = 2,842) sourced via yfinance (Yahoo).
• Dynamic selection: Dynamic Johansen cointegration is used to detect time-varying long-run equilibrium relationships; only intermittently or persistently cointegrated pairs are used for mean-reversion trading.
• Modeling: Spreads are modeled with two base learners — Deep Neural Network (DNN) and Long Short-Term Memory (LSTM) — and combined with a Dynamic Weighted Ensemble that updates weights based on recent predictive performance.
• Evaluation: Forecast accuracy assessed via MSE, RMSE, MAE, MAPE, MASE; uncertainty via 99% prediction intervals; trading performance assessed by directional accuracy, trading returns, drawdowns, and robustness under volatility regimes.
• Results: DWE achieves the lowest errors across metrics and more stable trading returns (lower drawdowns, better directional consistency) than static ensembles (bagging, boosting, stacking, voting). LSTM captures temporal persistence; DNN captures nonlinearities; the DWE adaptively prioritizes the best model over time.
• Deployment & compute: Training used high-performance computing; system deployed for real-time forecasting and signal generation on AWS (EC2 and Lambda).
• Assumptions and limits: Price series treated as non-stationary while combinations can be mean-reverting; sufficient liquidity assumed (transaction costs not explicitly modeled); rolling/sequential training used to avoid look-ahead bias.
• Practical caveats: Results rely on selected crypto pairs and period; transaction costs, market impact, and execution frictions were not explicitly incorporated.

Data & Methods

• Data: Daily close prices for BTC, ETH, LTC, XRP, BNB, USDT? (paper lists several; main empirical pairs are the six listed above) from 2018-01-02 to 2025-10-31 (2,842 obs), obtained via yfinance.
• Pair selection: Dynamic Johansen cointegration testing (time-varying formulation) to identify pairs whose long-run equilibrium relationship exists and/or changes over time.
• Spread construction: Mean-reverting spread derived from cointegrating vector(s) and used as target variable for forecasting.
• Forecast models:
  • DNN: multilayer feedforward network with ReLU activations; trained by SGD/backpropagation.
  • LSTM: recurrent architecture with input/forget/output gates to capture temporal dependencies and regime persistence.
  • Dynamic Weighted Ensemble (DWE): adaptive weighting scheme reallocating model weights in real time based on recent predictive performance and market state (versus fixed-weight ensembles).
• Training & evaluation: Rolling/sequential train/validation/test splits to prevent look-ahead bias; metrics = MSE, RMSE, MAE, MAPE, MASE; predictive uncertainty via 99% prediction intervals.
• Implementation: Trained on HPC resources; real-time inference and signal delivery on AWS EC2 and Lambda for automated on-demand forecasting.
• Trading rules: Trading signals derived from predicted spread deviations and prediction intervals; buy/sell rules generate market-neutral long/short positions with low-lag execution (assumed feasible under liquidity assumption).

Implications for AI Economics

• Modeling non-stationarity matters: Time-varying econometric structure (dynamic cointegration) combined with adaptive ML improves exploitability of statistical arbitrage opportunities in non-stationary markets such as crypto.
• Hybrid AI+econometrics is effective: Integrating economic structure (cointegration/mean reversion) with flexible AI learners (DNN/LSTM) and adaptive ensembles yields better out-of-sample performance than purely statistical or static-ensemble approaches.
• Operationalization requirements: Real-time AI-driven trading requires HPC/cloud infrastructure and robust rolling validation; firms deploying such systems must plan for continuous model monitoring and weight adaptation to regime shifts.
• Market efficiency and arms race: More adaptive, automated strategies can compress arbitrage profits and increase competition for short-lived inefficiencies, accelerating an “AI arms race” in high-frequency/quant trading.
• Risk and regulation: AI-enabled, rapidly adapting strategies can amplify volatility or propagate shocks if widely deployed; regulators and market designers should consider monitoring adaptive automated trading and incorporating liquidity/transaction cost effects into policy analysis.
• Research directions: Future AI-economics work should incorporate transaction costs, market impact, liquidity constraints, broader asset universes, robustness to adversarial regime shifts, causal interpretability, and stress-testing under extreme events to assess economic viability and systemic effects.

Suggested next steps for researchers/practitioners: - Add explicit transaction-cost and market-impact modeling to check profitability net of frictions. - Extend to larger cross-sections and intraday data to evaluate scalability and latency effects. - Compare alternative time-varying dependence models (nonlinear cointegration, state-space) and explore interpretability tools for AI forecasts. - Conduct multi-period backtests across distinct market regimes and run live paper-trading to validate execution assumptions.