The cryptocurrency market is highly non-linear, non-stationary and driven by sentiment volatility, which poses fundamental challenges for price prediction. The recently proposed Helformer model couples a Transformer-based architecture with Holt-Winters exponential smoothing and shows excellent performance for univariate forecasting. However, the model relies only on closing price data and does not incorporate external sentiment sources, and uses static decomposition parameters without taking into account market conditions. Finally, the Helformer model has only been evaluated on a single asset. To systematically address these limitations, this paper proposes the Multivariate Adaptive Helformer (MA-Helformer), an improved deep learning framework with three main contributions. First, the input feature space is enriched with a complete multivariate set of 13 technical indicators, six on-chain metrics, seven macroeconomic variables, and three sentiment scores from FinBERT. Second, an adaptive market regime detection module based on Hidden Markov Models (HMM) dynamically conditions Holt-Winters decomposition parameters on detected market state (Bull, Bear, Sideways). Third, a multi-horizon forecasting head is trained jointly to produce price forecasts for 1-day, 7-day, and 30-day horizons simultaneously. We experiment on Bitcoin (BTC), Ethereum (ETH) and Solana (SOL) using daily data from January 2017 to December 2024. MA-Helformer reduces next-day RMSE by 49.2% relative to the baseline Helformer, achieves a Sharpe Ratio of 2.34 in a realistic trading simulation that accounts for variable transaction costs and slippage, and shows particularly strong improvements in Bear market regimes. All reported performance gains are statistically significant ("p<0.01" ) according to the Diebold-Mariano test. Experimental results demonstrate that MA-Helformer consistently outperforms the baseline Helformer and several competitive forecasting models across multiple evaluation metrics and forecast horizons, while achieving statistically significant improvements under the adopted experimental protocol. Although the proposed framework demonstrates promising predictive performance, its evaluation is based on historical market data and simulated trading conditions; therefore, further validation using real-time deployment and additional financial assets remains future work.

Traditional sentiment analysis performs well in terms of polarity detection, but it cannot capture cognitive nuances needed for mental health monitoring. In this paper, we address two gaps in the current literature: the lack of task-driven guidance on how to choose transfer learning strategies and the lack of validation on cognitively rich da-tasets, other than standard polarity benchmarks. We present a systematic review of 150 transfer learning studies follow-ing PRISMA guidelines, summarising the findings into a decision framework that maps data availability, domain spec-ificity, and resource constraints to optimal model selection. We empirically validate our decision framework on a novel 4-class cognitive thought classification dataset with human-elicited samples and GPT-4 augmented data. Active learn-ing reduced annotation effort by 60% and achieved high inter-rater agreement (Fleiss’ Kappa = 0.82). According to our framework, 12 models were evaluated. BiLSTM-TF-IDF achieved 93.9% macro F1, outperforming transformer models by 0.5% with 9× less compute, supporting the medium-data/domain-specific path. Specifically, on our 4,944-sample domain-specific cognitive thought dataset using a single NVIDIA RTX 3090 GPU, BiLSTM-TF-IDF outperformed RoBERTa-base by 0.5% macro F1 (93.9% vs. 93.4%, McNemar p=0.004) via uncertainty-sampling active learning with SVM confidence threshold <0.40, while requiring approximately 9.4× less inference time and a 32× smaller model footprint (with a 6× lower peak GPU memory requirement). Uncertainty was quantified using least-confidence sam-pling (selecting samples where max class probability <0.40); the cross-entropy loss function was used for both the SVM-guided active learning and BiLSTM training. We also evaluate cross-dataset performance on IMDb and SST-2, demonstrating reasonable generalisability. Demographic analysis shows cognitive patterns that have implications for mental health. The dataset and code have been published at Zenodo (DOI: 10.5281/zenodo.17444289) to facilitate reproducible cognitive NLP research.