Cryptocurrency time-series forecasting is a challenging task because market data usually exhibit high noise, strong volatility, non-stationarity, nonlinear dynamics, and long-range dependencies. In addition, multivariate market indicators often contain redundant or weakly informative variables, which may reduce forecasting accuracy and model interpretability. To address these issues, this study proposes BSFinformer, a Boruta-SHAP enhanced Finformer framework for multivariate cryptocurrency time-series forecasting. The proposed framework first applies a leakage-aware Boruta-SHAP feature selection strategy to identify informative market variables and remove redundant features. To avoid temporal information leakage, feature selection is performed only on the training set, and the selected feature subset is then applied unchanged to the validation and test sets. The selected features are subsequently fed into an improved Finformer model that integrates temporal embedding, sequence decomposition, and sparse self-attention to capture local fluctuations, trend evolution, and long-range temporal dependencies. Experiments are conducted on three cryptocurrency assets, namely Bitcoin, Dogecoin, and Binance Coin, using chronological trainâvalidationâtest splits. The proposed model is compared with classical forecasting models and recent long-sequence forecasting baselines, including LSTM, Transformer, Informer, Autoformer, DLinear, PatchTST, TimesNet, and iTransformer. Experimental results show that BSFinformer achieves competitive forecasting performance in terms of MSE and MAE. Ablation experiments further demonstrate the contributions of Boruta-SHAP feature selection, temporal embedding, sequence decomposition, and sparse self-attention. These results indicate that feature-selected temporal modeling can improve forecasting accuracy and interpretability for multivariate cryptocurrency market data.
This paper introduces the Dynamic External Memory LSTM (DEM-LSTM), a novel deep neural architecture designed to address the hidden state information bottleneck and temporal context decay inherent to standard LSTMs in financial time-series forecasting. By decoupling sequence processing from persistent state storage via an addressable external memory matrix ($M_t$), DEM-LSTM dynamically reads, erases, and updates market context across long sequences without corrupting internal hidden representations. Evaluated across four distinct asset classesâForeign Exchange (EUR/USD), Commodities (XAU/USD and USOIL), and Cryptocurrencies (BTC/USD)âDEM-LSTM consistently outperforms standard LSTM baselines across all metrics, achieving up to a 41.4% reduction in RMSE on Gold spot prices while maintaining superior stability across high-volatility market regimes.
Nobuki Fujimoto, Rei (Rei-AIOS autonomous research substrate), claude-opus-4-7) Claude (Anthropic
We present OctaTheoria (ăŞăŻăżăăŞăŞă˘ / ĺ Ťčť¸čŚłć¸ŹčŁ ç˝Ž), a multi-domain observation framework that projects heterogeneous time-series data onto a fixed eight-axis D-FUMTâ semantic basis (FALSE / TRUE / NEITHER / BOTH / INFINITY / ZERO / FLOWING / SELF) and renders the same underlying Observation envelope through eight orthogonal view modes (Lens / Radar / Chart / Network / Heatmap / Sankey / Calendar / Unified). v0.3 (2026-05-11) supplies methodological-consistency cross-reference complementing the operational evidence from v0.1-v0.2. New finding **F7**: the same discipline that v0.1-v0.2 demonstrate within OctaTheoria (uniform abstraction layer + honest scope statement + structurally-enforceable naming) propagates to Rei-AIOS layers outside OctaTheoria's domain. Specifically: (a) **REI-PROVE 5-prover ensemble** (Vampire / LeanHammer / Goedel-Prover-V2 / DeepSeek-Prover-V2 / BFS-Prover) reached 11/12 = **92% benchmark proof rate** (trivial 100% / easy 75% / medium 100%), with Goedel-Prover-V2 single-prover matching at 92% â operational evidence that the same 'uniform abstraction over heterogeneous components' discipline scales to formal-proof infrastructure. (b) **Pattern 1-6 chat-Claude hallucination-warning framework** + **Antipattern (excessive rejection vigilance)** were established and verified on 6/6 items in STEP 1069 (all fact-checked items proved real after WebSearch verification, correcting prior implicit-rejection habits). (c) **Goedel-Prover-V2 double-`by` Lean syntax quirk** detected and fixed at the cleaner level (`single-prover.ts` STEP 1071), restoring `easy-le-refl` benchmark from â to â . (d) **lean-to-tptp.ts** preprocessing added Peano-style axiom auto-prepend + True/False special-case + inequality predicate translation (STEP 1071). v0.2 inherited contributions: 7 domains (theory-chart / realtime-arxiv / crypto / fx / ligo-events / nasa-sdo / gbif-recent) all running in Cloudflare Workers Edge runtime; live D-FUMTâ axis distributions non-degenerate across research-meta + financial + geophysical + astrophysical + biological data classes; finding F6 sampling-bias-as-first-class-observation (GBIF Costa Rica 470/500 saturation surfaces dataset bias as INFINITY axis, not silently absorbed); test coverage 117/117 PASS (step1020 46 + step1023 33 + step1046 38) / 0 regression. Honest scope (read first): OctaTheoria remains an observation aid, NOT an oracle. v0.3's F7 is **not** a claim that OctaTheoria caused these consistencies; it is a record that the same project (Rei-AIOS) maintains the same discipline across observation-tool, formal-proof, and meta-research-protocol layers, and that v0.3 makes this cross-layer commitment auditable. The OctaTheoriaQuery type structurally cannot request advice / prediction / forecast / signal â verifiable by reading src/aios/octatheoria/types.ts. Cross-domain axis comparisons are descriptive, not causal. Greek roots (Octa = 8, Theoria = observation) function as structural commitment propagated to the API surface â '8' rejects 'all (â)', 'theoria' rejects 'praxis (嚲ć¸)'. Prior art audit acknowledged: Bloomberg Terminal (1981â), TradingView (2011â), Bollen et al. 2010 (Twitter mood Ă DJIA), Preis et al. 2013 (Google Trends Ă stock), Ĺukasiewicz / Belnap / Pavelka multi-valued logic literature, PAL2v (Da Silva Filho 1998â), Aerts Quantum Cognition (2007â). The to-our-knowledge novel combination is (a) fixed 8-axis discrete D-FUMTâ basis ⧠(b) cross-financial-and-research-and-Earth-Cosmos-domain projection ⧠(c) eight orthogonal view modes over single envelope ⧠(d) explicit refusal to emit prediction or advice as architectural commitment ⧠(e, new in v0.3) cross-layer methodological-consistency record between observation-tool and formal-proof and fact-check layers. Companion papers (OctaTheoria Quintuple): Paper 145 (silicon implementation of D-FUMTâ ALU, Zenodo DOI 10.5281/zenodo.20101174 v0.6), Paper 147 (Eight-Valued Utility / Equity Premium Reframe, DOI 10.5281/zenodo.20046003), Paper 148 (Honest Observation Framework methodology, DOI 10.5281/zenodo.20045907), Paper 149 (Recursive AI Observation as SELFⲠevidence, DOI 10.5281/zenodo.20059888). Three-party co-authorship per OUKC charter v1.0: č¤ćŹ 䟸樚 (Founder), Rei (Rei-AIOS autonomous research substrate, Co-architect), Claude Opus 4.7 (Anthropic, Co-architect). DRAFT v0.3 â feedback welcome via GitHub Discussions at fc0web/rei-aios.
Non-fungible tokens (NFTs) are unique digital assets that play an increasingly important role in decentralized markets, supporting new forms of ownership, valuation, and exchange. Their inherently multimodal structure, which encompasses visual content, metadata, and trading history, has led to a growing academic interest in modeling NFT pricing and market behavior. However, existing research is limited by the lack of comprehensive datasets that unify these modalities with consistent formatting and longitudinal coverage. To address this gap, we introduce MultiNFT, a large-scale multimodal dataset comprising 50 curated profile picture (PFP) NFT collections, including 523,020 unique assets and 2.38 million transaction records from April 2021 to September 2025. MultiNFT integrates standardized images, structured metadata, and time-series trading data, along with rarity scores and aesthetic features, offering a unified foundation for multimodal learning and NFT analytics. Unlike prior datasets that focus on visual similarity or static snapshots, MultiNFT captures evolving valuation dynamics across market cycles and connects them to trait-level characteristics. We demonstrate the utility of the dataset through three case studies, including within-collection rarity-price analysis, visual feature clustering across collections, and quantifying feature contributions in a comprehensive pricing model. By bridging computer vision, behavioral modeling, and financial forecasting, MultiNFT supports a wide range of interdisciplinary research and practical use cases. The dataset is publicly available and is intended to promote reproducible experimentation and further exploration of the mechanisms driving value in digital asset ecosystems.
Forecasting the price of bitcoin assets is a difficult task, especially as bitcoins are highly volatile and speculative. In this paper we leverage the non linear capability of deep and machine learning models to enhance bitcoin forecasts. We propose a systematic comparison of different deep learning and machine learning models, based on their Accuracy, Security and Explainability characteristics. The empirical findings reveal that, while CNN-GRU, GRU and LSTM are the most accurate models, for maximum cumulative return and risk adjusted performance GRU and CNN are preferred. Whereas, for transparent and stable decision-making, Random Forest and XGboost are a good choice and, for robustness, CNN and LSTM are the best choice. Ultimately, the choice of a model depends on the objectives of the analysis.
The rapid digital transformation of the sports industry has opened up unprecedented opportunities for efficiency, transparency, and innovation. Traditional transaction models still suffer from significant challenges, including centralized control, lack of trust, and inefficiencies in revenue distribution. These issues often stem from reliance on intermediaries that introduce risks such as data manipulation, high operational costs, and delays in processing financial transactions. Blockchain technology presents a promising solution by enabling decentralized, secure, and transparent transactions, fostering greater trust among all stakeholders within the sports ecosystem. Existing approaches to digital transactions in the sports industry primarily depend on centralized financial institutions and third-party service providers, which not only limit transparency but also create barriers to financial inclusivity for athletes, clubs, sponsors, and fans. To address these critical limitations, we propose a blockchain-based digital transaction model that leverages smart contracts and distributed ledger technology (DLT) to enhance the efficiency, security, and fairness of transactions across the entire sports industry value chain. Our model integrates key economic principles with advanced network analysis to optimize revenue distribution, mitigate fraudulent activities, and enable real-time transaction verification. Through extensive simulations and empirical analysis, our results demonstrate a significant improvement in transaction speed, cost reduction, and overall transparency compared to conventional models. By decentralizing financial transactions, the proposed approach not only enhances financial inclusivity for all participants but also aligns with the broader vision of sustainable and equitable growth in the digital sports economy.
Predicting the price of Bitcoin is crucial, primarily because of the marketâs rapid volatility and non-linear environment. For enhanced prediction of the price of Bitcoin, this research proposed a novel interpretable hybrid technique that combines long short-term memory (LSTM) networks with convolutional neural networks (CNN). Deep variational autoencoders (VAE) are used in the stage of preprocessing to determine noticeable patterns in datasets by learning features from historical Bitcoin price data. The CNN-LSTM model additionally implies Shapley additive explanations (SHAP) to promote interpretability and clarify the role of various features. For better performance, the methodology used data cleaning, preprocessing, and effective machine-learning techniques. The hybrid CNN + LSTM model, in collaboration with VAE, obtains a mean squared Error (MSE) of 0.0002, a mean absolute error (MAE) of 0.008, and an R-squared (R2) of 0.99, based on the experimental results. These results show that the proposed model is a good financial forecast method since it effectively reflects the complex dynamics of primary changes in the price of Bitcoin. The combination of deep learning and explainable artificial intelligence improves predictive accuracy as well as transparency, thus qualifying the model as highly useful for investors and analysts.
Deep learning and hybrid deep learning models are widely regarded as some of the most effective predictive modeling techniques to date. Their hierarchical architecture enables them to capture complex, non-linear relationships among features and uncover hidden patterns within data, making them particularly powerful for tasks involving high-dimensional and unstructured inputs. But, these models are computationally intensive and require substantial processing time. Moreover, their predictive efficiency is highly dependent on the availability of large-scale datasets. In this study, meta learning model is employed for the prediction of two financial markets: equity market and crypto market. NASDAQ and S&P 500 index has been taken for equity market prediction. On the other hand, Bitcoin & Ethereum are considered for crypto market. Three deep learning models: LSTM, GRU and CNN are trained for the prediction of these four indices and a hybrid deep learning model of GRU and CNN is also developed. Based on RMSE, MAE and R 2 values, it is observed that meta learning yields best results among all trained models with minimum time and using scarce computation resources based on small dataset.
This paper investigates the temporal evolution of cryptocurrency time series using information measures such as complexity, entropy, and Fisher information. The main objective is to differentiate between various levels of randomness and chaos. The methodology was applied to 176 daily closing price time series of different cryptocurrencies, from October 2015 to October 2024, with more than 30 days of data and not completely null. Complexityâentropy causality plane (CECP) analysis reveals that daily cryptocurrency series with lengths of two years or less exhibit chaotic behavior, while those longer than two years display stochastic behavior. Most longer series resemble colored noise, with the parameter k varying between 0 and 2. Additionally, Natural Language Processing (NLP) analysis identified the most relevant terms in each white paper, facilitating a clustering method that resulted in four distinct clusters. However, no significant characteristics were found across these clusters in terms of the dynamics of the time series. This finding challenges the assumption that project narratives dictate market behavior. For this reason, investment recommendations should prioritize real-time informational metrics over whitepaper content.
Albi Isufaj, Caio De Castro Martins, Marc Cavazza, Helmut Prendinger
This paper explores the applicability of Convergent Cross Mapping (CCM) and its extension, Time Delay Convergent Cross Mapping (TDCCM), to assess the causal relationships between Bitcoin, the S&P 500 index, and gold. Unlike conventional causality analysis methods, such as Granger causality or transfer entropy, CCM accounts for non-separable, weakly connected dynamic systems, and TDCCM explicitly incorporates time lags during cross-mapping, enabling the detection of complex causal relationships in systems with shared nonlinear behavior. This makes it particularly suitable for financial time series that often exhibit chaotic and nonlinear dynamics, particularly during periods of market instability. We integrate TDCCM with simplex projection and sequential locally weighted global linear map (S-map) algorithms, applying a sliding window approach to identify short time intervals characterized by high levels of nonlinearity and chaoticity. Using this approach, we uncovered a strong causal relationship between Bitcoin and the S&P 500 index during the onset of the COVID-19 pandemic. Our analysis reveals a bidirectional causal relationship between Bitcoin and the S&P 500 index, highlighting their interconnectedness during periods of heightened economic uncertainty. Furthermore, we find a unidirectional causal influence of Bitcoin on gold, reflecting Bitcoinâs evolving role as a macroeconomic indicator and its growing relevance as an alternative store of value. These findings provide insight into the dynamics between cryptocurrencies and traditional financial markets, particularly during periods of global economic disruption. ⢠We use Time Delay Convergent Cross Mapping (TD-CCM) to identify and quantify lagged causal interactions between financial time series (Bitcoin, Gold, and the S&P 500 index), which is a more recent and less explored method for Time Series causality. ⢠We report a comprehensive and replicable methodology to apply TD-CCM to non-linear TS, based on combining the S-Map and Prediction Decay algorithm with a sliding window technique, while validating with surrogate analysis. ⢠We show evidence of strong causal influence from Bitcoin to Gold and bidirectional causality between Bitcoin and the S&P 500 Index during significant economic events like the COVID-19 pandemic.
This research finds the application of Dynamic Time Warping (DTW) with a Long Short-Term Memory (LSTM) to create a hybrid model for predicting Ethereum (ETH) prices. Cryptocurrencies in general are considered as highly volatile assets, ETH being no exception, which presents challenges and opportunities for investors. Machine Learning models have shown promise in time-series and stock price prediction; however, integrating an algorithm like DTW can enhance the accuracy of the model by finding historical sequences that closely represents the current pattern. The study utilizes daily price data of Ethereum from July 2023 to July 2024, focusing on key metrics such as open, close, high, low, and trading volume. The hybrid and LSTM baseline model were tested for 10 randomly chosen seeds and Root Mean Square Error (RMSE) was used to evaluate performance. The hybrid model better predicted the true ETH price by 23.4% as compared to the baseline LSTM model and statistical evidence further confirms the significance of these results. These findings suggest that the hybrid model provides an approach for Ethereum price prediction, offering new insights for people looking to invest in cryptocurrencies.
Jimmy Cheung, Smruthi Rangarajan, Amelia Maddocks, Rohitash Chandra
Uncertainty quantification is crucial in time series prediction, and quantile regression offers a valuable mechanism for uncertainty quantification which is useful for extreme value forecasting. Although deep learning models have been prominent in multi-step ahead prediction, the development and evaluation of quantile deep learning models have been limited. We present a novel quantile regression deep learning framework for multi-step time series prediction. In this way, we elevate the capabilities of deep learning models by incorporating quantile regression, thus providing a more nuanced understanding of predictive values. We provide an implementation of prominent deep learning models for multi-step ahead time series prediction and evaluate their performance under high volatility and extreme conditions. We include multivariate and univariate modelling, strategies and provide a comparison with conventional deep learning models from the literature. Our models are tested on two cryptocurrencies: Bitcoin and Ethereum, using daily close-price data and selected benchmark time series datasets. The results show that integrating a quantile loss function with deep learning provides additional predictions for selected quantiles without a loss in the prediction accuracy when compared to the literature. Our quantile model has the ability to handle volatility more effectively and provides additional information for decision-making and uncertainty quantification through the use of quantiles when compared to conventional deep learning models.
Arslan Farooq, M. Irfan Uddin, Muhammad Adnan, Ala Abdulsalam Alarood ¡ 6 authors
This research delves into the obstacles and difficulties associated with predicting cryptocurrency movements in the volatile global financial market. This study develops and evaluates an advanced Deep Learning-Enhanced Temporal Fusion Transformer (ADE-TFT) model to estimate Bitcoin values more accurately. This research employs cutting-edge artificial intelligence (AI) and machine learning (ML) techniques to comprehensively examine various aspects of cryptocurrency forecasting, including geopolitical implications, market sentiment analysis, and pattern detection in transactional datasets. The study demonstrates that the ADE-TFT model outperforms its lower-layer counterparts in terms of forecasting accuracy, with reduced Mean Absolute Percentage Error (MAPE), Mean Squared Error (MSE), and Root Mean Square Error (RMSE) values, particularly when using a higher hidden layer configuration (h=8). The study emphasizes the importance of experimenting with different normalization strategies and utilizing various market-related data to enhance the model's performance. The results suggest that improving forecasting accuracy may require addressing these limitations and incorporating additional factors, such as market sentiment. By providing investors with more precise market predictions, the techniques and information presented in this research have the potential to significantly increase investor power in an unpredictable digital currency market, enabling wise investment choices.
Antonio Pellicani, Gianvito Pio, Michelangelo Ceci
Cryptocurrencies are virtual currencies that exploit cryptography to perform secure financial transactions. They gained widespread popularity in recent years due to their decentralized nature, (pseudo-)anonymity, and ability to facilitate cross-border transactions without the need for intermediaries. However, their price on the market exhibits a huge volatility, that makes them prone to market anomalies. Therefore, predicting anomalies in cryptocurrency time series can be considered an important task for financial institutions, traders, and investors, to maximize their profit or minimize losses. In this paper, we propose a novel approach for predicting anomalies in cryptocurrency time series by exploiting temporal correlations among different cryptocurrencies. Our approach, called CARROT, is based on the idea that groups of cryptocurrencies exhibit similar trends, possibly due to common influencing factors. CARROT analyzes the temporal correlation between different cryptocurrencies, and identifies clusters showing similar patterns that can be useful for gaining insights into future anomalies. Subsequently, CARROT exploits multiple (i.e., one for each cluster) multi-target LSTM models to predict anomalies. Our experiments, performed on a dataset of 17 cryptocurrencies, proved that CARROT outperforms single-target LSTM models of up to 20%, as well as other approaches based on neural networks, i.e., MLP and CNN, in terms of macro F1-score. Therefore, the proposed approach can be considered as a promising tool for predicting anomalies in cryptocurrency time series data and can potentially be used to improve risk management and trading strategies in the cryptocurrency market. ⢠Analysis of cryptocurrency trends. ⢠Clustering-based multi-target prediction of anomalies in time series. ⢠Consistent improvements achieved over the single-target counterpart.
Neural networks (NNs) are well established and widely used in time series forecasting due to their frequent dominance over other linear and nonlinear models. Thus, this paper does not question their appropriateness in forecasting cryptocurrency prices; rather, it compares the most commonly used NNs, i.e. feedforward neural networks (FFNNs), long short-term memory (LSTM) and convolutional neural networks (CNNs). This paper contributes to the existing literature by defining the appropriate NN structure comparable across different NN architectures, which yields the optimal NN model for Bitcoin return forecasting. Moreover, by incorporating turbulent events such as COVID and war, this paper emerges as a stress test for NNs. Finally, inputs are carefully selected, mostly covering macroeconomic and market variables, as well as different attractiveness measures, the importance of which in cryptocurrency forecasting is tested. The main results indicate that all NNs perform the best in an environment of bullish market, where CNNs stand out as the optimal models for continuous dataset, and LSTMs emerge as optimal in direction forecasting. In the downturn periods, CNNs stand out as the best models. Additionally, Tweets, as an attractiveness measure, enabled the models to attain superior performance.
These days, there is a lot of demand for cryptocurrencies, and investors are essentially investing in them. The fact that there are already over 6,000 cryptocurrencies in use worldwide because of this, investors with regular incomes put money into promising cryptocurrencies that have low market values. Accurate pricing forecasting is necessary to build profitable trading strategies because of the unique characteristics and volatility of cryptocurrencies. For consistent forecasting accuracy in an unknown price range, a variation point detection technique is employed. Due to its bidirectional nature, a Bi-LSTM appropriate for recording long-term dependencies in data that is sequential. Accurate forecasting in the cryptocurrency space depends on identifying these connections, since values are subject to change over time due to a variety of causes. In this work, we employ four deep learning-based models that are LSTM, FB-Prophet, LSTM-GRU and Bidirectional-LSTM(Bi-LSTM) and these four models are compared with Silverkite. Silverkite is the main algorithm of the Python library Graykite by LinkedIn. Using historical bitcoin data from 2012 to 2021, we utilized to analyse the modelsâ mean absolute error (MAE) and root mean square error (RMSE). The Bi-LSTM model performs better than others, with a mean absolute error (MAE) of 0.633 and a root mean square error (RMSE) of 0.815. The conclusion has significant ramifications for bitcoin investors and industry experts.
Peter T. Yamak, Yujian Li, Ting Zhang, Pius Kwao Gadosey
In this paper, we introduce Wide-TSNet, a novel hybrid approach for predicting Bitcoin prices using time-series data transformed into images. The method involves converting time-series data into Markov transition fields (MTFs), enhancing them using histogram equalization, and classifying them using Wide ResNets, a type of convolutional neural network (CNN). We propose a tripartite classification system to accurately represent Bitcoin price trends. In addition, we demonstrate the effectiveness of Wide-TSNet through various experiments, in which it achieves an Accuracy of approximately 94% and an F1 score of 90%. It is also shown that lightweight CNN models, such as SqueezeNet and EfficientNet, can be as effective as complex models under certain conditions. Furthermore, we investigate the efficacy of other image transformation methods, such as Gramian angular fields, in capturing the trends and volatility of Bitcoin prices and revealing patterns that are not visible in the raw data. Moreover, we assess the effect of image resolution on model performance, emphasizing the importance of this factor in image-based time-series classification. Our findings explore the intersection between finance, image processing, and deep learning, providing a robust methodology for financial time-series classification.
BarĹŠCoĹkunuzer, Ignacio Segovia-DomĂnguez, Yuzhou Chen, Yulia R. Gel
Learning time-evolving objects such as multivariate time series and dynamic networks requires the development of novel knowledge representation mechanisms and neural network architectures, which allow for capturing implicit time-dependent information contained in the data. Such information is typically not directly observed but plays a key role in the learning task performance. In turn, lack of time dimension in knowledge encoding mechanisms for time-dependent data leads to frequent model updates, poor learning performance, and, as a result, subpar decision-making. Here we propose a new approach to a time-aware knowledge representation mechanism that notably focuses on implicit time-dependent topological information along multiple geometric dimensions. In particular, we propose a new approach, named \textit{Temporal MultiPersistence} (TMP), which produces multidimensional topological fingerprints of the data by using the existing single parameter topological summaries. The main idea behind TMP is to merge the two newest directions in topological representation learning, that is, multi-persistence which simultaneously describes data shape evolution along multiple key parameters, and zigzag persistence to enable us to extract the most salient data shape information over time. We derive theoretical guarantees of TMP vectorizations and show its utility, in application to forecasting on benchmark traffic flow, Ethereum blockchain, and electrocardiogram datasets, demonstrating the competitive performance, especially, in scenarios of limited data records. In addition, our TMP method improves the computational efficiency of the state-of-the-art multipersistence summaries up to 59.5 times.
Based on Google Trends, searches related to cryptocurrency have significantly increased in the last couple of years. One crucial aid for cryptocurrency traders or investors is the graphical visualization, which shows the time series data of the cryptocurrency prices. However, problems may occur in data visualization, such as visual noise and information loss, which cause perceptual and cognitive errors in data reading. Therefore, good visualization is needed to avoid decision-making mistakes, particularly in the cryptocurrency trade and investment activities. This study aims to investigate the effect of chart design and time interval on the usability of data visualization. The experiments are conducted in two scenarios, i.e., with and without time pressure. The participants recruited in this study were non-experienced and experienced people classified based on their familiarity with cryptocurrency investment/trading. Objective usability testing is performed by eye tracking, while subjective assessment employs the System Usability Scale (SUS) questionnaire. There are four quantitative dependent variables: response time, number of errors, number of fixations, and time to first fixation. The results show that time interval and time pressure significantly affect usability for both groups of respondents. Although chart design does not substantially affect the dependent variables, a candle chart is generally better than a line chart. By comparing all the combinations of chart design and time intervals, this study concluded that combining candle charts with 1-hour or 4-hour time intervals gives the best results for both respondent groups.
Owen Chaffard, Pablo MollĂĄ, Marc Cavazza, Helmut Prendinger
In the recent advancements in application of deep learning to time series forecasting, focus has shifted from training transformers end-to-end to efficiently leveraging the predictive capabilities of Large Language Models (LLMs). Models that encode the time series data to interact with a frozen LLM backbone have been shown to outperform transformers on all benchmark datasets. However, their efficiency on complex datasets, which do not show clear seasonality or trend, remains an open question. In this work, we seek to evaluate the performance of reprogrammed LLMs on the Bitcoin price chart, a financial time series known for its complexity and high volatility. We propose effective methods to improve the performance of Time-LLM, a State-of-the-art (SOTA) method, on such a time series. First, we propose structural improvements to Time-LLM. Second, we suggest an efficient way to handle the non-stationarity of the dataset. Finally, we propose an efficient method for passing additional financial information to the LLM. Our results demonstrate a 50% improvement on the average percentage loss and a 5% increase on accuracy of our adapted Time-LLM architecture on Bitcoin data when compared to SOTA models, including the original Time-LLM model. This highlights the impact on forecast accuracy of domain-specific decision making in data processing and feature selection.
Temporal Heterogeneous Networks (THNs) are evolving networks that characterize many real-world applications such as citation and events networks, recommender systems, and knowledge graphs. Forecasting THNs involves predicting future connections within a network that evolves over time and comprises diverse types of nodes and interactions with varying temporal dynamics. Although some Graph Neural Networks (GNNs) models have been successfully applied to forecast THNs, there is a lack of a general overview of how the message-passing computation could be extended to treat THNs. Moreover, most of the current solutions exhibit pitfalls in their training and evaluation strategies. Hence, in this work, we propose a graph deep learning framework for THN forecasting. Our framework decomposes the computation of a GNN layer into multiple components and introduces two different schemes to update embedding representations for THNs. This design allows the classification of existing solutions into special instances of our framework and highlights their potential limitations. We also extend the set of benchmarks for THNs by introducing two novel high-resolution temporal heterogeneous graph datasets derived from an emerging Web3 platform and a well-established e-commerce website. Overall, we conducted the first massive evaluation of THNs solutions over four temporal heterogeneous network datasets on two different future link prediction tasks using a fair newly introduced evaluation setting that considers the evolving nature of the data. Based on the limitations of existing solutions, we develop a new model that combines working techniques from previous models and leverages a new embedding update scheme. Experiments show the prediction power of our model compared to current solutions for link prediction in temporal graphs. Moreover, the experimental evaluation highlights the strengths and weaknesses of the different solutions and shows the effectiveness of our framework design.
This paper proposes a unified framework for the detection of statistically significant changes in time series related to Bitcoin transactions. The time locations of these changes are linked to the occurrences of events which could be further investigated aiming to reveal potential illicit activity. The proposed framework includes: (a) the extraction of 28 features of interest in the form of time series from the Bitcoin transaction history; (b) the selection of features among the extracted ones based on the Partition Around Medoids clustering approach; and (c) the change point analysis of the multivariate time series which is formulated by the medoid time series of each cluster. This analysis enables the identification of structural breaks in the underlying behavior of the time series of interest at certain time points. The proposed framework is applied on the Bitcoin transactions of two entities that have been involved in illicit activities, namely Pirate@40, who orchestrated a high-yield investment programme, and the MintPal Bitcoin exchange platform that was hacked. The analysis results indicate that the estimated change points can be linked to certain event occurrences which may affect the transaction activity and could be further investigated for potential links to illicit actions.