A Hybrid SVR-Based Framework for Cryptocurrency Price Forecasting and Strategy Backtesting

ABSTRACT Cryptocurrency price forecasting has gained increasing attention due to the market’s high volatility and structural complexity. While many recent studies have explored deep learning architectures, including attention- and transformer-based models, existing research still faces notable limitations: (i) inconsistent feature engineering choices, (ii) limited examination of hybrid machine-learning models, and (iii) a lack of transparent trading evaluation using realistic backtesting assumptions. To address these gaps, this study develops a hybrid forecasting and trading framework based on Support Vector Regression (SVR) combined with a set of rule-based technical strategies. Using four major cryptocurrencies – BTC, ETH, XRP, and LTC – from 2018 to 2020, the proposed framework integrates thirteen technical indicators with a sliding-window scheme and compares SVR against Random Forest (RF) and Long Short-Term Memory (LSTM) benchmarks. Empirical results show that SVR offers a competitive balance between predictive accuracy and computational efficiency, particularly in moderate-volatility regimes. The strategy backtesting further demonstrates that SVR-driven signals can outperform traditional technical rules under certain market conditions, although limitations remain for highly volatile assets such as Bitcoin. The study contributes to the literature by clarifying feature-design choices, evaluating SVR within a multi-asset setting, and providing reproducible code and datasets through an open-access repository. Introduction The rapid rise of cryptocurrencies has significantly reshaped the global financial landscape, attracting increasing interest from individual investors, institutional traders, and researchers alike. As decentralized digital assets, cryptocurrencies such as Bitcoin (BTC), Ethereum (ETH), Ripple (XRP), and Litecoin (LTC) offer new opportunities for high-return investment, but are also characterized by extreme price volatility, speculative market behavior, and limited regulatory oversight. These challenges make accurate price forecasting both essential and difficult, especially when compared to traditional financial instruments. Given the limitations of conventional forecasting models in capturing the nonlinear and rapidly changing dynamics of cryptocurrency markets, researchers have increasingly turned to machine learning techniques for enhanced prediction performance. Among these, Support Vector Regression (SVR) has proven effective for time-series forecasting due to its robustness in handling small, noisy datasets and its ability to model complex relationships (Cortes and Vapnik Citation1995; Ding et al. Citation2021). However, accurate price prediction alone does not necessarily translate into profitable investment outcomes. To address this gap, recent studies have combined forecasting models with trading strategies, aiming to enhance investment decision-making through data-driven methods (Fister et al. Citation2019; Jaquart, Köpke, and Weinhardt Citation2022; Liu et al. Citation2023; Liu, Tsyvinski, and Wu Citation2022). This study proposes a hybrid framework that integrates SVR-based cryptocurrency price forecasting with four rule-based technical trading strategies: buy-and-hold, Bollinger Bands, moving average crossover, and a combined KD-MACD indicator strategy. Using historical data from 2018 to 2020, we examine the predictive performance of SVR and evaluate the effectiveness of each trading strategy through backtesting. The objective is to determine whether integrating machine learning models with technical indicators can improve investment returns and offer practical insights into cryptocurrency trading behavior. By exploring both prediction accuracy and strategy profitability, this work contributes to the growing body of literature on intelligent financial systems and their applications in emerging markets (Chiang et al. Citation2012; Liu et al. Citation2023; Liu, Tsyvinski, and Wu Citation2022; Tingjia Citation2021; X. Wu et al. Citation2020). In addition to evaluating prediction accuracy and backtesting results, this study contributes to the literature in two key ways. First, it bridges machine learning forecasting with practical investment strategies across multiple cryptocurrencies, thereby emphasizing both predictive performance and real-world trading applicability. Second, we conduct a comparative analysis between SVR, LSTM, and random forest models, highlighting SVR’s balance between computational efficiency and prediction accuracy. These aspects address gaps in prior studies that often focus exclusively on either model design or trading rules, but not their integration. As such, this work provides a comprehensive and application-driven perspective to cryptocurrency forecasting. Cryptocurrency forecasting has attracted significant research attention, and recent studies have increasingly adopted advanced deep learning approaches – including attention-based and transformer-based architectures (Lee Citation2025; Lian Citation2024; Rafi et al. Citation2024) – to capture complex temporal dependencies. These models demonstrate strong predictive capabilities and reflect a major shift in modern sequence-modeling research. Despite this progress, the broader literature still exhibits several structural limitations. First, many studies lack transparency in feature-engineering choices or provide limited justification for how selected indicators relate to underlying market dynamics. Second, evaluation protocols often differ across works, with inconsistent backtesting assumptions that make the practical effectiveness of forecasting models difficult to compare. Third, most studies focus on a single asset or a single family of algorithms, leaving limited understanding of how different models generalize across cryptocurrencies with distinct volatility characteristics. Finally, reproducibility remains an ongoing concern, as many works do not provide open-source implementations, standardized datasets, or clear documentation of hyperparameters. To address these gaps, this study makes the following contributions: Development of a hybrid SVR-based forecasting framework, integrating thirteen technical indicators with a sliding-window scheme and applying it across four major cryptocurrencies (BTC, ETH, XRP, LTC). Transparent methodological design, including explicit justification of feature selection, hyperparameter settings, and model configuration decisions to enhance interpretability and reproducibility. Comprehensive multi-model evaluation, comparing SVR with Random Forest and LSTM benchmarks, and analyzing algorithmic strengths, limitations, and cross-asset generalizability. Realistic trading evaluation, incorporating rule-based strategies and clarifying the influence of assumptions such as transaction costs and volatility regimes on overall performance. A fully reproducible research package, offering open-access datasets, Jupyter notebooks, and hyperparameter logs through a publicly available Zenodo repository. These contributions collectively strengthen the methodological rigor, practical relevance, and reproducibility of machine-learning-based cryptocurrency forecasting. Related work The search for usable crypto price prediction strategies is as new as the crypto industry itself. There is a large literature on predicting stock prices, but the number of studies on cryptocurrency price forecasting is very small due to the young age of crypto and the absence of a fully developed regulatory framework. A review of the stock price prediction literature reveals three primary approaches: Research and analysis to determine the intrinsic value of companies, considering factors such as industry trends, financial statements, and other verifiable data (Heaton and Lucas Citation1999). Sentiment analyses that track the flow of funds and attempt to verify trends being followed by major traders in order to ride their coattails to make successful investments. Technical analyses that use historical data to predict future trends, with prices and volumes frequently serving as key indicators (Brock, Lakonishok, and LeBaron Citation1992). The first two analytical approaches are considered challenging due to the difficulty of obtaining complete data sets, therefore researchers are emphasizing a mix of technical indicators and machine learning processes to assess the potential for crypto profitability. For example, Tingjia (Tingjia Citation2021) combined Granger causality and self-vector regression models to study degree of investor attention in Taiwan for five cryptocurrencies (Bitcoin, Litecoin, Ethereum, Tether and Ripple) and five comprehensive stock indexes (Taiwan Weighted Stock Price, Shanghai Composite, Hang Seng, Nikkei and Korea Composite). Tether had the strongest correlations with returns from the Taiwan Weighted Stock Price and Hang Seng Indexes, and Bitcoin had a significantly positive correlation with returns from the Shanghai Stock Exchange Index. According to Yukun Liu, Tsyvinski, and Wu (Citation2022), specific cryptocurrency market characteristics, market size, and market momentum show positive potential for accurately predicting cryptocurrency returns. They analyzed nine factors associated with stock market price predictions that exerted significantly positive effects on long- and short-term trading strategies, and then found positive correlations between all of them and three factors: cryptocurrency market characteristics, size, and momentum. Caporale, Gil-Alana, and Plastun (Citation2018) used R/S analysis and fractional integration to examine cryptocurrency market persistence in Bitcoin, Litecoin, Ripple and Dash transactions between 2013 and 2017, and reported a high degree of persistence in cryptocurrency markets over time, thus giving support to the use of trend-based trading strategies to improve crypto transaction profitability. A number of researchers have used machine learning models in their efforts to predict cryptocurrency market trends, thereby providing a wealth of empirical evidence associated with market efficiency theory and statistical arbitrage strategies (Jaquart, Köpke, and Weinhardt Citation2022; Yujun; Liu et al. Citation2023). Individual research teams have used recurrent neural network (RNN), long short-term memory (LSTM), gated recurrent unit (GRU), random forest (RF), and gradient boosting (GB) machine models, among others, to predict intraday crypto price movements and performance. According to most of these efforts, machine learning models excel at capturing the nonlinear characteristics of cryptocurrency markets and their connections with price movements. For example, Yujun Liu et al. (Citation2023) used LSTM and GRU models to demonstrate how risk-adjusted returns exceed market benchmarks in statistical arbitrage strategies, and to illustrate how the overall crypto market challenges a weak form of market efficiency. Jaquart, Köpke, and Weinhardt (Citation2022) compared ordinary least squares (OLS) and extreme gradient boosting (XGB) models, and found that the XGB model performed better in predicting cryptocurrency returns, especially in scenarios involving exceptionally low or high risk. Combined, these and other studies underscore the value of applying machine learning models to cryptocurrency markets in terms of prediction accuracy and statistical arbitrage strategy returns. Building upon the growing literature on machine learning-based forecasting, several traditional models have also been tested for cryptocurrency prediction. Traditional machine learning approaches such as logistic regression and random forest have been applied to cryptocurrency price prediction. For instance, Kim et al. (Citation2016) used user sentiment data and machine learning models to forecast short-term price movements, emphasizing the impact of market sentiment on volatility. Jiang and Liang (Citation2017) proposed a deep reinforcement learning framework for cryptocurrency portfolio management, showcasing the adaptability of dynamic learning environments. However, these studies largely emphasize predictive accuracy or theoretical modeling without incorporating trading simulations or backtesting. In contrast, our study contributes by integrating a predictive model – support vector regression (SVR) – with multiple rule-based trading strategies, thereby bridging the gap between forecasting and practical investment decision-making. This application-driven design enhances both the interpretability and real-world relevance of machine learning in cryptocurrency trading. Aside from machine learning-based approaches, time series models have also been widely explored. Other models used to predict crypto price trends include ARIMA, neural networks, and exponential moving average (EMA) (Karasu et al. Citation2018; Mills Citation2019; Mudassir et al. Citation2020; Roy, Nanjiba, and Chakrabarty Citation2018; Yenidoğan et al. Citation2018). They used a combination of the VAR model, impulse response function, and variance decomposition to explore relationships among both Bitcoin and gold markets and the Taiwan stock index. They found that among these prediction models, ARIMA is the best model. They also reported weak correlations among the three measures that were insufficient for explaining their respective long-term changes. The list of approaches that have been used to study stock price changes and to examine trading strategies includes Bollinger Bands, the Relative Strength Index (RSI), and the Moving Average Convergence and Divergence (MACD) indicator, as discussed in prior studies (Fister et al. Citation2019; Takeuchi and Lee Citation2013; Wei Citation2021; X. Wu et al. Citation2020; Yang et al. Citation2020). There is disagreement on how to interpret the findings of these methods, with some describing a Bollinger basic strategy as having the best performance, others preferring RSI or the Parabolic SAR indicator, and still others emphasizing the strengths of active over passive trading strategies. In a study involving high-risk versus risk-free assets, Chiang et al. (Citation2012) described stochastic advantage theory as particularly useful in helping investors choose optimal asset allocations. Based on our review of strategies regularly applied to stock investment scenarios, for the present study we limited our scope to technical indicators without fundamental or chip information. We believe this decision fits with significantly different market mechanisms but similar technical indicators for stocks and cryptocurrencies. In the final section of this paper we compare these strategies in terms of price prediction effectiveness. Recent studies have increasingly applied hybrid deep learning architectures to cryptocurrency forecasting. For instance, Livieris et al. (Citation2021) proposed a CNN-LSTM model that effectively captures both spatial and temporal dependencies in cryptocurrency time series, demonstrating improved predictive accuracy compared to standalone models. From a broader perspective, Otabek and Choi (Citation2024) provided a comprehensive review of forecasting and trading strategies in cryptocurrency markets, encompassing a wide range of machine learning approaches, technical indicators, and performance metrics. Building upon these prior works, our study explores a hybrid SVR-based framework integrated with rule-based technical trading strategies. Benefiting from the foundations laid by previous research, which has largely emphasized predictive accuracy or model sophistication, our approach seeks to bridge forecasting with practical trading applications through empirical backtesting. This contributes an application-oriented evaluation across multiple cryptocurrencies, aiming to achieve a balance between model interpretability, computational efficiency, and trading performance. Recent years have witnessed rapid progress in the application of deep learning architectures to cryptocurrency forecasting. Traditional recurrent neural networks and LSTM-based models have been widely utilized to capture sequential patterns in price movements. However, the emergence of attention mechanisms and transformer-based architectures has further advanced the forecasting capability of deep learning models. For example, Helformer model (Kehinde et al. Citation2025) introduced an attention-based framework specifically tailored for cryptocurrency price forecasting, demonstrating strong predictive performance through adaptive weighting of temporal information. Similarly, Rafi et al. (Citation2024) incorporated transformer-based models with on-chain metrics to improve volatility forecasting accuracy, highlighting the importance of integrating blockchain-specific signals. Recent comparative analyses in the literature indicate that transformer-based architectures often capture long-term temporal dependencies more effectively than conventional recurrent neural networks, reflecting a broader shift toward attention-driven models in cryptocurrency forecasting. Beyond individual-model innovations, multi-asset and hybrid designs have also been explored. Lee (Citation2025) proposed a Temporal Fusion Transformer-based approach that integrates both technical indicators and on-chain variables for multi-cryptocurrency trading strategies. Their results highlight the potential benefits of combining attention mechanisms with domain-specific financial features. Meanwhile, ensemble and hybrid ML – DL frameworks remain a parallel research direction, offering interpretable and computationally efficient alternatives for practitioners working with moderate-sized datasets. Overall, while transformer-based and attention-driven architectures represent important recent developments, the literature contains relatively fewer studies examining the comparative performance and interpretability of traditional machine-learning methods such as Support Vector Regression (SVR) within multi-asset settings. This gap underscores the need for transparent feature design, reproducible experimental setups, and realistic trading evaluations – areas addressed by the present study. Models Research structure and data The four cryptocurrencies of focus for this study were Bitcoin (BTC), Ethereum (ETH), Ripple (XRP) and Litecoin (LTC). All crypto information employed in this research was scraped from Yahoo Finance using the yfinance API. Daily closing prices over a three-year period from January 1, 2018 to December 31, 2020 served as the primary dataset. The thirteen features shown in served as input data; feature descriptions and calculation methods are explained in later sections. Data for 1/1/18 to 6/31/20 were used for model training, data for 7/1/20 to 12/31/20 for model testing. For performance evaluations, the SVR model from Python’s Scikit-Learn package was used to establish price predictions to be integrated with the trading strategies described in Trading Strategies. Market value rankings were used to select cryptocurrencies with longer durations. Acknowledging that many cryptocurrencies with large market capitalizations have only recently emerged, we needed to make adjustments to overcome training data deficiencies that might interfere with SVR performance. Sliding windows were employed to improve machine prediction accuracy and to reduce prediction errors. We then used the evaluation model to compare prediction results. Technical indicators were combined with the four trading strategies, prediction results were applied to perform price backtesting, and final return rates were compared with actual prices to evaluate trading strategy performance. To address potential data quality issues and market biases commonly found in cryptocurrency datasets, we applied several preprocessing steps. Specifically, we used min-max normalization to scale all input features into a 0–1 range, which helps reduce the influence of outliers and improves model convergence – particularly for SVR and LSTM models. We also ensured that all selected cryptocurrencies (Bitcoin, Ethereum, Ripple, and Litecoin) had complete and continuous historical data from 2018 to 2020. The dataset was obtained from Yahoo Finance, which aggregates pricing information across major exchanges and performs internal smoothing to reduce minor discrepancies. Nevertheless, we acknowledge that some residual noise and potential market manipulation patterns (e.g., pump-and-dump schemes) may still exist, which remains a limitation of this study. The study period from January 2018 to December 2020 covers multiple cryptocurrency market phases, including post-bubble correction, prolonged consolidation, and the bullish recovery of late 2020. This ensures that both upward and downward price dynamics are represented in the training and testing windows. The data were not extended beyond 2020 to maintain consistency and reproducibility, as later periods involve heterogeneous data sources and extreme regime shifts. Furthermore, to examine the effect of different market conditions, the backtesting was conducted in two sub-periods (Backtesting Data for Two Time Segments), corresponding to moderate-volatility and high-volatility segments. Machine model SVM, a supervised learning algorithm, was first proposed by Corinna Cortes and Vladimir Vapnik in 1993 (Cortes and Vapnik Citation1995). It is generally used for problems involving small samples, nonlinearity, and high dimensions. Initially applied to binary classification problems, SVM has a reputation as an efficient model for research in the fields of medicine and industry, among others. Given a set of training data and a linear classifier, each data point belongs to one of two categories. Classifiers may be straight lines, curves, or planes. Classification lines are identified based on the specific algorithms and conditions involved, with support vectors for individual lines requiring maximization. Very few types of real-world data are linearly classifiable. For low-dimension and linearly inseparable data, one of four kernel methods (liner, polynomial, radial basis function, sigmoid) can be used to project limited-dimension spaces into higher-dimension spaces, thus making the data linearly separable. For the present study we employed the radial basis function kernel method, the most commonly used kernel function in SVM classification research (Ding et al. Citation2021). Support vector regressions (SVRs), an SVM extension, were specifically designed for regression problems. According to the SVR model, a loss function parameter ε indicates a tolerable prediction error value along the f(x) regression line. The larger the ε value, the lower the model accuracy; smaller ε values result in greater accuracy, but also increase the potential for overfitting. Furthermore, the hyperparameter settings used in the SVR model are described as follows to enhance the reproducibility of the study. The SVR model was implemented with a radial basis function (RBF) kernel to effectively capture nonlinear relationships in financial time series. The key hyperparameters were set as: - : This penalty parameter controls the trade-off between model complexity and training error. A larger C allows the model to fit the training data more closely but increases the risk of overfitting. - : A very small kernel coefficient that determines the influence range of each support vector, enabling the model to generalize better by smoothing the decision surface. The epsilon parameter retained its default value of 0.1, providing a margin of tolerance for prediction errors. These settings were selected based on common practices and adjusted empirically for our dataset. For completeness, the hyperparameters C, , and were tuned through a limited grid-search procedure within the training window only, exploring , , and . The parameter set that minimized the mean-squared error on the validation portion of the training data was adopted. This deterministic search was chosen for its transparency and reproducibility, providing stable results without the additional randomness inherent in meta-heuristic methods such as PSO or GA. In addition to SVR, this study includes Random Forest and LSTM as representative baseline models from ensemble learning and deep sequential architectures. These models were selected to provide a balanced comparison across distinct algorithmic paradigms while maintaining a transparent and reproducible experimental design. More advanced architectures such as transformer-based models typically require larger datasets, extensive hyperparameter tuning, and greater computational complexity, and thus fall beyond the methodological scope of the present study. Evaluating SVR and its baselines under a controlled and interpretable setting aligns with the objective of examining the performance and applicability of a hybrid SVR-driven forecasting framework. Besides, SVR has strong performance on moderate-sized and noisy financial datasets, as well as its robustness and computational efficiency. Technical Indicators and Sliding Windows The following six technical indicators (eigenvalues) were used for SVR machine learning and for comparing trading strategies: Moving average (MA) represents an average price for a defined time period. Commonly used MA durations are 5, 10, 20, 60, 120 and 240 days. An increase in MA means an increase in price. MA values are calculated using the formula (1) (1)Stochastic oscillator (commonly referred to as a KD indicator) is a momentum analysis method first proposed by George Lane in 1950 (Lai, Chen, and Caraka Citation2019; M. Wu and Diao Citation2015). KD is used to determine whether a market is too hot or too cold. Indicators consist of K value and D value lines. K values are also referred to as “fast lines,” with greater sensitivity to price fluctuations. In comparison, D value lines (“slow lines”) respond less quickly and with less sensitivity. Stochastic oscillator calculations begin with a raw stochastic value expressed as (2) (2) where n denotes trading time, Cn the closing price on the nth day, Hn the highest price during the past n days, and Ln the lowest price during the past n days. Daily K and D values are calculated as with α generally set to 1/3. In the absence of K or D values for the previous day, the α value is set to 0.5. 3. Bollinger Bands (BBands), a technical analysis method created by John Bollinger in 1980, are mainly used to predict future price trends. They consist of 20-day moving averages, two standard deviations, and three rails (upper, lower and middle). In the first BBand step, the value of the middle rail is calculated using MA with an N value of 20: Next, the MA standard deviation (σ) is calculated as followed by upper and lower rail calculations: 4. Moving average convergence/divergence (MACD) is a commonly used stock trading analysis tool created by Gerald Appel in the 1970 s. After examining the intensity, direction, energy, and trend cycles of stock price changes, it reviews stock price support and pressure information to determine the best timing for buying and selling a particular stock. The most commonly employed MACD time durations are 12 or 26 days for fast lines and 9 days for slow lines. The first of four calculation steps involves exponential moving average (EMA), with t denoting the current date. The second step consists of calculating differential lines (DIFs) for 12- and 26-day time periods: Next, the MACD value (also known as the DEM value) is calculated for a 9-day period: Last, DEM is subtracted from DIF, with the result used to draw a histogram/MACD bar. 5. The relative strength index (RSI), created by Welles Wilder in 1978, is a momentum oscillator that measures price movement speed and change. Its calculation process involves two steps, the first focused on rising and falling price averages: The second involves calculating actual RSI: 6. Williams %R (also known as W%R) was introduced by stock market investor Larry Williams in his 1973 trade edition, How I Made One Million Dollars Last Year Trading Commodities. It is a price oscillator that uses highest and lowest prices in the most recent prior period to estimate recent price trends. W%R has many similarities with the KD indicator K-line. Its basic formula is with Cn the closing price on the nth day, Hn the highest price during the past n days, and Ln the lowest price during the past n days. To avoid model training errors resulting from large variations in feature values, normalization was applied to compress all features to [0 ~ 1], using the formula where Xmax denotes maximum value and Xmin minimum value. Sliding windows of two different lengths were used for data collection and group experiments, with each window addressing both featured and targeted data. After capturing a data set, the window in question was moved forward one day to capture the next set (Ho Citation1995). In this study, the five-day sliding window was applied for data collected between 1/1/18 and 6/31/20 (used for model training), and the one-day window was applied for data collected between 7/1/20 and 12/31/20 (used for testing purposes). The five-day training window was selected based on exploratory testing showing that shorter windows helped the SVR model better capture rapid market fluctuations common in cryptocurrency markets, while the one-day prediction horizon reflects a practical short-term forecasting scenario widely adopted in prior forecasting studies. To clarify the feature-selection rationale, the thirteen technical indicators were chosen to represent three complementary dimensions of market behavior: trend (MA, KD, MACD), momentum (RSI, Williams %R), and volatility (Bollinger Bands). These indicators have been widely adopted in prior stock and cryptocurrency forecasting research. In line with ensemble-based feature-importance analyses discussed in previous studies, volatility- and momentum-related indicators are generally expected to contribute most strongly to predictive performance. To prevent overfitting, the SVR model was trained with a fixed RBF kernel and hyperparameters tuned solely on the training window using a limited grid search, coupled with rolling-window validation to ensure temporal independence. Trading Strategies Four trading strategies were selected for performance comparisons Buy and hold Virtual currency was purchased on day 1 of the SVR data testing period (7/1/20) and held until the final day (12/31/20), regardless of changes in price or any other factor. Bollinger Band indicator-based strategy involving five signals The first “buy signal” is a price line passing through the lower Bollinger Band rail from bottom to top, the second a price line passing through the middle rail from bottom to top, and the third a price line situated between the middle and upper rails that indicates a long position – that is, a higher value than the previous day. An investor purchases cryptocurrency whenever the market matches any of these signals. The first “sell signal” is a price line that crosses the lower Bollinger Band rail from top to bottom, and the second a price line between the middle and lower rails. The second price line indicates a short position – a lower value than the previous day. The investor sells cryptocurrency whenever the market matches either one of these conditions. Moving average indicator-based strategy Three moving averages were examined: MA5, MA10 and MA20, with cryptocurrency purchases triggered by a MA5 > MA10 > MA20 signal on any day, and a MA10 > MA5 signal the preceding day. Accordingly, the MA5 and MA10 moving averages show golden crosses with MA5 climbing upward. Selling is triggered by a MA5 < MA10 < MA20 signal on any individual day and MA10 < MA5 the preceding day. In this case the MA5 and MA10 moving averages resemble a death cross, with MA5 trending downward. A combined KD-MACD-based strategy An investor makes a purchase when the two KD lines make a golden cross and the MACD bar is positive. The combination of a KD death cross and negative MACD bar is a signal to sell. The buying-and-selling time period for each strategy was the same as for the SVR algorithm test period (7/1/20–12/31/20). Predicted prices were used for transactions and actual prices for performance comparisons. None of the trades were leveraged, eliminating the possibility of inflated investment volumes. Results Accuracy and Prediction Comparisons Actual and SVR-predicted price trends for Bitcoin, Ethereum, Ripple and Litecoin are respectively shown in . The Bitcoin price curves do not intersect in because the data for the second half of 2020 indicate that the actual price exceeded the predicted price, with the distance between the two lines increasing in step with the speed of the actual price increase. In comparison, the two Ethereum lines shown in crisscross at roughly the same rate throughout the graph. The lines for Litecoin () and Ripple () also intertwine at multiple points, but at irregular intervals; where actual price changes were relatively flat, predicted prices fluctuated more widely, and where actual prices show wide fluctuation, predicted prices were roughly similar. According to the data shown in , the largest errors in predicted prices were for Bitcoin – significantly larger than for the other three cryptocurrencies. The lowest error rate/greatest accuracy was observed for Ethereum (Mean Absolute Percentage Error [MAPE] = 0.07). MSE: mean squared error; MAE, mean absolute error; MAPE, mean absolute percentage error. Trading Strategy Comparison The best buy-and-hold return rate was for Bitcoin, followed by Ethereum, Litecoin and Ripple, in that order (). For the Bollinger Band indicator-based strategy, Ethereum showed the best return rate, followed by Bitcoin, Ripple and Litecoin (). According to , the average number of transactions for each currency was approximately 150. present data for the moving average indicator-based strategy. Ripple had the best return rate (121%), followed by Bitcoin and Ethereum; a negative return rate was observed for Litecoin. According to , fewer than 10 transactions were recorded for each of the four cryptocurrencies during the data testing period. For the combined KD-MACD-based strategy, the data shown in indicate the highest return rate (75%) for Ripple, followed by Litecoin and Bitcoin; a negative return rate was observed for Ethereum. According to , there were approximately 20 transactions for each of the four cryptocurrencies during the data testing period. The mixed profitability observed across strategies can be attributed to inherent market-regime sensitivities. Trend-following indicators such as moving-average crossovers tend to underperform in sideways or noisy markets due to frequent whipsaw signals, whereas mean-reversion indicators like Bollinger Bands become less effective during sustained high-volatility expansions. Similarly, momentum-based rules such as KD – MACD may lag during rapid reversals, contributing to the negative returns seen in certain cryptocurrencies. These results indicate that strategy underperformance is strongly linked to structural market conditions rather than deficiencies in the SVR forecasts themselves. Although the main focus of this study is to evaluate the profitability of rule-based trading strategies driven by SVR predictions, we acknowledge that risk management plays a critical role in real-world trading systems. In the current implementation, trading signals are generated based on technical indicators such as Bollinger Bands, Moving Averages, and MACD. However, no explicit stop-loss mechanisms, dynamic position sizing, or capital exposure constraints were incorporated. This lack of embedded risk control may limit the practical applicability of the strategies under extreme market conditions. Future research could enhance the realism of the trading framework by integrating common risk management techniques, such as volatility-adjusted position sizing, Kelly criterion-based allocation, or dynamic trailing stop strategies. These additions could help mitigate drawdowns and improve long-term portfolio stability. To clarify the backtesting assumptions, all simulated trades were executed using daily closing prices without leverage, transaction costs, or slippage. These simplifications help isolate the predictive contribution of the SVR-based forecasts but may overstate absolute returns compared with real-world execution. Assuming a modest transaction fee (approximately 0.1–0.5%) would proportionally reduce total profits while maintaining the same performance ranking among strategies. Furthermore, no dynamic position sizing, stop-loss, or capital allocation rules were applied in this version. Future research will incorporate these elements – such as volatility-adjusted position sizing and dynamic stop-loss mechanisms – to enhance the realism and robustness of the trading framework. Backtesting Data for Two Time Segments Some exceptional increases and decreases occurred during the data collection, model training, and testing periods. We therefore established two time segments for price backtesting, a process that uses historical data to simulate investment or trading strategies for purposes of evaluating strategy performance in different markets or settings. Price increases and decreases ranged from small to moderate during the four-month segment 1 (7/1/20–10/31/20), and were generally large during the two-month segment 2 (11/1/20–12/31/20). Price change data for each strategy are presented in . They indicate clear connections between the highest returns and largest price movements for the most successful strategy, and the largest losses for the least successful, especially during the second segment – for example, a 4.9% loss for the Bollinger Band strategy (). Prediction Model Comparison and Performance Evaluation In addition to machine-learning baselines (RF and LSTM), we briefly reference classical time-series models such as ARIMA and GARCH as representative linear and volatility-driven benchmarks in prior literature. While not re-estimated here, they help contextualize our SVR-based results within traditional econometric frameworks. Furthermore, all experiments follow a rolling-origin validation to prevent look-ahead bias and ensure robust out-of-sample evaluation. To better understand the effectiveness and trade-offs of our proposed SVR-based approach, we conducted additional experiments comparing SVR with two widely used models: Random Forest (RF) and Long Short-Term Memory (LSTM). The models were trained and tested using identical datasets and feature sets to ensure consistency in evaluation. The comparison, summarized in , shows that while LSTM achieved the lowest error metrics in many cases, it requires significantly greater computational resources, longer training times, and access to GPU-based hardware, which may not be feasible for all users or applications. On the other hand, RF, although relatively lightweight, exhibited a critical weakness in forecasting sudden upward trends, particularly visible in the Bitcoin price chart (), where it failed to respond adequately during steep climbs. In contrast, SVR offers a well-balanced trade-off between computational efficiency and predictive accuracy. It handles nonlinear patterns with reasonable precision while remaining computationally accessible. Therefore, despite not being the top performer in every metric, SVR stands out as a practical and reliable solution for cryptocurrency forecasting in real-world scenarios. Extreme volatility and sudden upward jumps in Bitcoin prices further exacerbate absolute forecast errors, as models based on historical patterns – whether SVR, RF, or LSTM – struggle to anticipate abrupt nonlinear movements. In the broader context of existing studies, recent deep learning – based approaches such as LSTM-, CNN-, and transformer-based architectures have often demonstrated strong predictive performance when larger datasets or complex temporal dependencies are present. Compared with these works, our SVR-based framework prioritizes interpretability, computational efficiency, and reproducibility; therefore, the comparison remains conceptual rather than numerical due to differences in datasets, feature sets, and forecasting horizons. For completeness, we provide the key parameter settings used for the RF and LSTM models. The Random Forest (RF) model was implemented using n_estimators = 100 to ensure stable predictions, with a fixed random_state = 42 for reproducibility. Other parameters were left at their default values. The LSTM model was constructed with one hidden layer of 50 memory units (units = 50) and used the Adam optimizer with the mean squared error (MSE) loss function. The model was trained for 20 epochs with a batch size of 32. Input sequences were reshaped to (samples, time steps, features), where the time steps matched the number of input features. The architecture was intentionally kept simple to balance performance and training time. While the comparative metrics in show consistent performance ranking across models, future studies could formally verify these differences through statistical tests such as the Diebold – Mariano (DM) test for pairwise forecast comparison or the Friedman test for multi-model evaluation. Because the current test window spans only six months of out-of-sample data, applying such tests may not yield robust statistical power. Moreover, because standard SVR does not generate probabilistic forecasts, the present study does not include formal uncertainty quantification. Meaningful uncertainty estimation would require methodological extensions – such as bootstrap resampling or quantile-based modeling – which lie beyond the scope of this work but represent an important direction for future research. Although each cryptocurrency model in this study was trained and tested independently, this design reflects the distinct volatility and behavioral patterns observed in different digital assets. Preliminary testing of a model trained on BTC and applied to ETH indicated that direct cross-asset transferability results in substantially higher prediction errors, suggesting that asset-specific tuning remains important. Future work may extend this framework by exploring multi-asset or transfer-learning approaches – such as transfer SVR, domain adaptation, or multi-task learning architectures – to capture shared dynamics among correlated cryptocurrencies and enhance model generalizability. Finally, we acknowledge that cryptocurrency markets evolved considerably after 2020, particularly throughout the pandemic and the 2021–2024 boom-and-bust cycles. Validating the proposed framework on these more recent periods represents a promising direction for future research, allowing for assessment of robustness under different volatility regimes and structural market conditions. Conclusion This study introduces a hybrid SVR-based framework for forecasting cryptocurrency prices and evaluating investment performance through multiple trading strategies. The proposed model demonstrates strong predictive capabilities, particularly for Ethereum and Ripple, under volatile market conditions. Backtesting results indicate that while a buy-and-hold strategy yields the highest overall returns in trending markets, technical rule-based strategies such as moving average and KD-MACD combinations can offer competitive returns with more dynamic risk control. Our findings suggest that integrating machine learning models with technical trading strategies can enhance investment decision-making by capturing both nonlinear price patterns and market behavior. As cryptocurrency markets continue to mature and evolve, this framework provides a foundation for future research that explores alternative algorithms (e.g., LSTM, ensemble methods), real-time predictive mechanisms, and adaptive trading systems tailored to investor risk profiles. Overall, the approach highlights the practical potential of AI-driven forecasting models in supporting intelligent financial strategies. In future work, we plan to extend this research in several directions. First, additional input variables such as market sentiment, trading volume, and macroeconomic indicators could be incorporated to enrich model inputs and improve predictive performance. Second, advanced deep learning models such as Transformer-based architectures or hybrid attention mechanisms may further enhance forecasting accuracy. Third, more realistic portfolio simulations could be designed by embedding risk control mechanisms, including position sizing, volatility-based stop-loss rules, and drawdown limits. Finally, reinforcement learning frameworks may be explored to dynamically adjust trading decisions in response to SVR-based forecasts and changing market conditions. In summary, the novelty of this study lies in its integration of machine learning-based forecasting with rule-based trading strategies across multiple cryptocurrencies. Unlike many prior works that focus solely on predictive modeling or investment simulation in isolation, our framework bridges both domains to evaluate real-world profitability. The cross-model comparison further reinforces the practical value of SVR, balancing predictive performance and computational feasibility. This comprehensive and application-oriented approach adds empirical depth to the growing body of research on AI-driven financial decision support. Disclosure statement No potential conflict of interest was reported by the author(s). Data availability statement The data that support the findings of this study are openly available in Zenodo at https://doi.org/10.5281/zenodo.15261521., reference number (Wang and Huang Citation2025). Additional information Funding References - Brock, W., J. Lakonishok, and B. LeBaron. 1992. “Simple Technical Trading Rules and the Stochastic Properties of Stock Returns.” The Journal of Finance 47 (5): 1731–23. https://doi.org/10.1111/j.1540-6261.1992.tb04681.x. - Caporale, G. M., L. Gil-Alana, and A. Plastun. 2018. “Persistence in the Cryptocurrency Market.” Research in International Business and Finance 46:141–148. https://doi.org/10.1016/j.ribaf.2018.01.002. - Chiang, Y.-C., M.-C. Ke, T. L. Liao, and C.-D. Wang. 2012. “Are Technical Trading Strategies Still Profitable? Evidence from the Taiwan Stock Index Futures Market.” Applied Financial Economics 22 (12): 955–965. https://doi.org/10.1080/09603107.2011.631893. - Cortes, C., and V. Vapnik. 1995. “Support-Vector Networks.” Machine Learning 20 (3): 273–297. https://doi.org/10.1023/A:1022627411411. - Ding, X., J. Liu, F. Yang, and J. Cao. 2021. “Random Radial Basis Function Kernel-Based Support Vector Machine.” Journal of the Franklin Institute 358 (18): 10121–10140. https://doi.org/10.1016/j.jfranklin.2021.10.005. - Fister, D., J. Mun, V. Jagrič, and T. Jagrič. 2019. “Deep Learning for Stock Market Trading: A Superior Trading Strategy?” Neural Network World 29 (3): 151–171. https://doi.org/10.14311/NNW.2019.29.011. - Heaton, J., and D. Lucas. 1999. “Stock Prices and Fundamentals.” NBER Macroeconomics Annual 14:213–242. https://doi.org/10.1086/654387. - Ho, T. K. 1995. “Random Decision Forests.” Paper Presented at the Proceedings of 3rd International Conference on Document Analysis and Recognition. - Jaquart, P., S. Köpke, and C. Weinhardt. 2022. “Machine Learning for Cryptocurrency Market Prediction and Trading.” The Journal of Finance and Data Science 8:331–352. https://doi.org/10.1016/j.jfds.2022.12.001. - Jiang, Z., and J. Liang. 2017. “Cryptocurrency Portfolio Management with Deep Reinforcement Learning.” Paper Presented at the 2017 Intelligent Systems Conference (IntelliSys). - Karasu, S., A. Altan, Z. Saraç, and R. Hacioğlu. 2018. “Prediction of Bitcoin Prices with Machine Learning Methods Using Time Series Data.” Paper Presented at the 2018 26th Signal Processing and Communications Applications Conference (SIU). - Kehinde, T., O. J. Adedokun, A. Joseph, K. M. Kabirat, H. A. Akano, and O. A. Olanrewaju. 2025. “Helformer: An Attention-Based Deep Learning Model for Cryptocurrency Price Forecasting.” Journal of Big Data 12 (1): 81. https://doi.org/10.1186/s40537-025-01135-4. - Kim, Y. B., J. G. Kim, W. Kim, J. H. Im, T. H. Kim, S. J. Kang, and C. H. Kim. 2016. “Predicting Fluctuations in Cryptocurrency Transactions Based on User Comments and Replies.” PLOS ONE 11 (8): e0161197. https://doi.org/10.1371/journal.pone.0161197. - Lai, C. Y., R.-C. Chen, and R. E. Caraka. 2019. “Prediction Stock Price Based on Different Index Factors Using LSTM.” Paper Presented at the 2019 International Conference on Machine Learning and Cybernetics (ICMLC). - Lee, M. C. 2025. “Temporal Fusion Transformer-Based Trading Strategy for Multi-Crypto Assets Using On-Chain and Technical Indicators.” Systems 13 (6): 474. https://doi.org/10.3390/systems13060474. - Lian, J. 2024. “Comparative Analysis of LSTM, GRU and Transformer Deep Learning Models for Cryptocurrency ZEC Price Prediction Performance.” Paper Presented at the 9th International Conference on Financial Innovation and Economic Development (ICFIED 2024). - Liu, Y., Z. Li, R. Nekhili, and J. Sultan. 2023. “Forecasting Cryptocurrency Returns With Machine Learning.” Research in International Business and Finance 64:101905. https://doi.org/10.1016/j.ribaf.2023.101905. - Liu, Y., A. Tsyvinski, and X. Wu. 2022. “Common Risk Factors in Cryptocurrency.” The Journal of Finance 77 (2): 1133–1177. https://doi.org/10.1111/jofi.13119. - Livieris, I. E., N. Kiriakidou, S. Stavroyiannis, and P. Pintelas. 2021. “An Advanced CNN-LSTM Model for Cryptocurrency Forecasting.” Electronics 10 (3): 287. https://doi.org/10.3390/electronics10030287. - Mills, T. C. 2019. Applied Time Series Analysis: A Practical Guide to Modeling and Forecasting. Academic press. - Mudassir, M., S. Bennbaia, D. Unal, and M. Hammoudeh. 2020. “Time-Series Forecasting of Bitcoin Prices Using High-Dimensional Features: A Machine Learning Approach.” Neural Computing and Applications 37 (28): 1–15. https://doi.org/10.1007/s00521-020-05129-6. - Otabek, S., and J. Choi. 2024. “From Prediction to Profit: A Comprehensive Review of Cryptocurrency Trading Strategies and Price Forecasting Techniques.” IEEE Access 12:87039–87064. https://doi.org/10.1109/ACCESS.2024.3417449. - Rafi, M. A., S. I. Shaboj, I. Rasul, M. K. Miah, M. R. Islam, and A. Ahmed. 2024. “Cryptocurrency Volatility Forecasting Using Transformer-Based Deep Learning Models and On-Chain Metrics.” Journal of Economics, Finance and Accounting Studies 6 (1): 119–130. - Roy, S., S. Nanjiba, and A. Chakrabarty. 2018. “Bitcoin Price Forecasting Using Time Series Analysis.” Paper presented at the 2018 21st International Conference of Computer and Information Technology (ICCIT), Dhaka, Bangladesh. - Takeuchi, L., and Y.-Y. A. Lee. 2013. Applying Deep Learning to Enhance Momentum Trading Strategies in Stocks Technical Report. CA, USA: Stanford University Stanford. - Tingjia, B. 2021. “Can Cryptocurrency Attention Predict the Return of a Weighted Stock Price Index. ( Master’s thesis, Fu Jen Catholic University. - Wang, S.-W., and C.-Y. Huang. 2025. “Dataset for: A Hybrid SVR-Based Framework for Cryptocurrency Price Forecasting and Strategy Backtesting.” https://doi.org/10.5281/zenodo.15261521. - Wei, C. R. 2021. “Empirical Study of Bollinger Bands Trading Strategies. In-Service Master’s Program Master’s Thesis, National Kaohsiung University of Science and Technology. - Wu, M., and X. Diao. 2015. “Technical Analysis of Three Stock Oscillators Testing MACD, RSI and KDJ Rules in SH & SZ Stock Markets.” Paper presented at the 2015 4th International Conference on Computer Science and Network Technology (ICCSNT), Harbin, China. - Wu, X., H. Chen, J. Wang, L. Troiano, V. Loia, and H. Fujita. 2020. “Adaptive Stock Trading Strategies with Deep Reinforcement Learning Methods.” Information Sciences 538:142–158. https://doi.org/10.1016/j.ins.2020.05.066. - Yang, H., X.-Y. Liu, S. Zhong, and A. Walid. 2020. “Deep Reinforcement Learning for Automated Stock Trading: An Ensemble Strategy.” Paper presented at the Proceedings of the first ACM international conference on AI in finance. - Yenidoğan, I., A. Çayir, O. Kozan, T. Dağ, and Ç. Arslan. 2018. “Bitcoin Forecasting Using ARIMA and Prophet.” Paper presented at the 2018 3rd international conference on computer science and engineering (UBMK), Sarajevo, Bosnia and Herzegovina.