Identifying Moments of Decision Making on Trade in Financial Time Series Using Fuzzy Cluster Analysis

This article is a summary of the scientific work published by Vladyslav Kabachii, Roman Maslii, Serhii Kozlovskyi, Oleksandr Dronchack

The original source can be accessed via DOI 10.33111/nfmte.2023.175

The pursuit of consistent, long-term profitability is central to trading strategy development, and achieving this goal depends heavily on a trader's ability to recognize shifts in market trends with speed and accuracy. This necessity underscores the importance of not only forecasting future prices but also strategically identifying key trading decision points—specifically, the optimal moments to enter (Buy) and exit (Sell) the market. This concept, referred to in the article as the “problem of identifying trading decision points,” forms the core of effective decision-making in financial time series analysis.

Traditional time series forecasting methods such as Simple Moving Averages, ARMA, and ARIMA models have been widely employed in economic forecasting but are often ill-suited for trading. These models generally focus on one-step-ahead predictions, which may be adequate in macroeconomic contexts but fall short in real-time trading, where the key to profitability lies in timing entries and exits, managing risk, and reacting to market volatility. The article argues that merely predicting the next price point is insufficient; instead, models must focus on price direction and the broader context of market dynamics to guide actionable trading decisions.

Empirical studies show that even sophisticated forecasting models optimized using least squares methods deliver limited predictive power in terms of directional accuracy—only slightly above 50%. One proposed solution has been to predict the direction of price movement rather than its exact value, using fuzzy logic models built on rule bases informed by Elliott Wave Theory. An alternative and perhaps more pragmatic approach is to rely on technical analysis to identify the most favorable points of entry and exit, taking into account various indicators such as trend strength, price volatility, support and resistance levels, and momentum oscillators.

As demonstrated in Figure 1, which visualizes Buy and Sell decisions for the EUR/USD pair on an hourly chart (H1), the effectiveness of a trading strategy is tied to the precise timing of these decisions. In practice, each point on the chart usually reflects the hourly Close price, though Open, High, Low, and Close prices are all part of the dataset. This reinforces the complexity of financial time series, where small changes can have large implications over short timeframes.

The article also critiques the limited utility of classical technical analysis, noting that many strategies offer marginal advantages, with directional accuracy rarely exceeding 55-60%. To enhance this, traders often adopt multi-timeframe analysis—for instance, using a higher timeframe like H4 as a trend filter for signals generated on a lower timeframe like M15. This filtering helps eliminate false signals that run counter to the dominant market trend, improving overall strategy reliability.

Beyond classical methods, the article emphasizes the growing role of intelligent systems—such as neural networks, fuzzy logic, and genetic algorithms—in financial forecasting. These tools excel at modeling complex nonlinear relationships and integrating numerous variables, making them well-suited to the dynamic nature of financial markets. However, the effective use of such technologies requires not only advanced data science skills but also a deep understanding of trading environments, making the blend of data science expertise and trading acumen increasingly essential.

The article argues for a shift in focus from narrow price prediction toward a broader, more strategic understanding of market behavior. By combining advanced intelligent systems with domain-specific knowledge and multi-timeframe analysis, traders can significantly enhance their ability to identify profitable decision points—transforming data into action and analysis into opportunity.

From Patterns to Profits: How Fuzzy Clustering Transforms Trading Decision-Making

Modern trading success hinges less on predicting exact future prices and more on accurately identifying recognizable patterns that indicate when to enter or exit the market. These “trading decision points” are the key moments that determine profitability, especially in fast-moving or volatile conditions. Traditional models that forecast one-step-ahead price movements often fall short because they don’t account for broader market behavior or provide actionable insights for real-world trading decisions.

To address this, analysts have increasingly shifted toward pattern recognition—identifying trends and behaviors within financial time series that typically precede profitable buy or sell opportunities. Early approaches used rule-based systems developed from trader intuition and hands-on market experience. While effective, these models lacked flexibility and scalability. Today, fuzzy logic models offer a more adaptive alternative by enabling nuanced interpretation of uncertain data, making them ideal for capturing market signals that don’t follow rigid rules.

Another major advancement in this field is the application of clustering techniques. Clustering allows the grouping of similar market patterns or investor behaviors, revealing underlying structures in price data without requiring pre-labeled outcomes. Among various algorithms, fuzzy clustering—especially fuzzy C-means—stands out as particularly suited to financial data. Unlike traditional methods that force each data point into a single group, fuzzy clustering allows patterns to belong to multiple clusters to varying degrees. This better reflects the real-world uncertainty and overlapping trends seen in financial markets.

Clustering has already proven useful in tasks like segmenting investor behavior, identifying trading anomalies, and reducing noise in data. However, many existing models still fall short of delivering precise entry and exit signals. The true business value lies not just in classifying trends, but in turning those classifications into real-time trading actions.

That’s why the proposed approach focuses on integrating fuzzy clustering with technical indicators to identify trading signals based on the degree of similarity between current and historical market patterns. The goal is to recognize not only when trends are forming or ending, but also when those patterns suggest a strong probability of profitable action. These insights can feed directly into automated trading systems, enhancing decision-making speed, reducing emotional bias, and improving the consistency of outcomes.

By applying this methodology, financial institutions, trading platforms, and even individual traders can elevate their strategy from reactive to predictive—gaining a competitive edge in complex and uncertain markets.

Designing a Fuzzy Clustering-Based Trading System: A Step-by-Step Methodological Framework

To design a trading system capable of identifying optimal market entry and exit points, this methodology uses a fuzzy clustering approach—specifically, the Fuzzy C-Means (FCM) algorithm. The process is divided into structured, sequential stages, each essential for turning raw financial data into actionable trading insights. These steps are visually represented in Figure 2.

Stages of the proposed methodological approach

1. Data Preparation
The process begins with the collection of relevant historical market data for selected financial instruments and timeframes. This includes open, close, high, and low prices for each interval, serving as the foundation for subsequent analysis.

2. Feature Engineering
In this stage, technical indicators and custom features are computed—such as moving averages, momentum oscillators, and volatility measures. These enhance the data and help reveal underlying patterns in market behavior.

3. Standardisation
To ensure comparability across features, all values are transformed into standard scores (z-scores), normalizing the dataset. This prevents any single feature from disproportionately influencing the clustering results.

4. FCM Clustering
The normalized training dataset is processed using the Fuzzy C-Means (FCM) algorithm. Each data point is assigned a degree of membership to multiple clusters, reflecting the ambiguous and overlapping nature of real market states.

5. Determining Whether Each Point Belongs to Each Cluster of Test Data
The trained FCM model is applied to new (test) data. For each new point, the model calculates its degrees of membership to the pre-identified clusters, enabling pattern recognition on unseen data.

6. Creating a Trading Strategy
Based on membership values, trading signals are generated. If a data point's membership to a Buy or Sell cluster exceeds a certain threshold, a corresponding trading action is initiated. These rules translate cluster memberships into actionable strategy components.

7. Modelling, Results, Optimisation
The system is backtested on historical data to evaluate trading performance. Metrics such as profit, drawdown, and accuracy are analyzed to refine thresholds, cluster parameters, and indicator settings—ensuring the model’s robustness and practical effectiveness.

Choosing the Right Metrics to Evaluate Trading System Performance

In the world of financial trading, success is not defined by how precisely a model forecasts a future price, but by how effectively it helps traders make profitable decisions. Traditional statistical metrics like Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) are commonly used in time series forecasting. However, while these are valuable in academic or engineering contexts, they are often misleading or irrelevant in trading environments.

Why? Because these metrics are designed to assess prediction accuracy by measuring deviations between forecasted and actual values. They do not account for whether those predictions translate into profitable trades or effective risk management. For instance, a model might predict the next price point with high precision but miss a critical turning point in the market, resulting in poorly timed buy or sell signals. In practice, this could lead to losses—despite the statistical model showing “good performance.”

Trading is fundamentally about decision-making under uncertainty, where the key challenge is not just predicting what will happen, but knowing when to act. The financial outcome of a trade—whether a profit is realized or capital is lost—depends heavily on the timing of entry and exit decisions. As such, models must be evaluated on metrics that align with financial objectives, not just mathematical accuracy.

To properly assess a trading system’s performance, a more suitable set of financially oriented metrics is needed. These include:

Why These Metrics Matter

Unlike abstract accuracy measures, these four metrics capture the essence of real-world trading: making money while managing risk. They allow traders, analysts, and fund managers to evaluate strategies not only for their return potential but also for their stability, resilience, and suitability under different market conditions.

In today’s fast-paced and algorithm-driven markets, where automated systems can execute thousands of trades in minutes, using appropriate performance metrics is critical. Relying solely on statistical accuracy can lead to the false belief that a model is effective, when in fact it might be systematically misaligned with market behavior.

Therefore, when building, testing, and deploying trading systems—especially those powered by advanced techniques such as machine learning or fuzzy logic—it is imperative to adopt financially grounded evaluation criteria. These offer a more comprehensive, realistic, and actionable view of how a model performs in practice.

Data Preparation and Feature Engineering for Clustering-Based Trading Systems

To build a reliable and data-driven trading model, the foundation lies in structured, high-quality historical data. In this study, currency pairs were selected due to their global liquidity, availability of long-term data, and relevance to both retail and institutional traders. The dataset includes four major forex pairs—EUR/USD, AUD/USD, USD/CAD, and USD/JPY—analyzed across two timeframes: Daily (D) and Four-Hour (H4). For each instrument, 800 data points were collected per timeframe, covering distinct periods for training and testing, enabling robust model evaluation.

Historical data was sourced using the MetaTrader 4 platform and exported in a standardized CSV format. The data preprocessing and feature engineering stages were conducted in Python using Jupyter Notebook—an ideal environment for combining computation, visualization, and analysis.

A critical aspect of the modeling process is the extraction of meaningful features from raw financial time series. Rather than relying on raw price data alone, the study leverages technical analysis indicators, which are mathematical functions designed to detect trends, reversals, and momentum shifts. Specifically, these widely respected indicators were used:

These features (MACD1, MACD2, S%K, and S%D) were selected based on practical trading experience and refined through empirical testing. Importantly, while the original closing price data (Close) is retained for calculating indicators, it is not used directly in the clustering process, ensuring the model focuses solely on derived signals rather than market noise.

To standardize the input data for clustering and ensure numerical consistency, all feature values were normalized using z-score transformation. This allows the clustering algorithm to treat each indicator with equal weight.

‍

‍
‍

This data preparation and feature engineering pipeline forms the backbone of the fuzzy clustering methodology. By carefully selecting only the most informative indicators and decoupling them from raw price inputs, the model gains a clearer understanding of behavioral market patterns—paving the way for precise entry and exit signals based on market structure rather than price noise.

The results of the modelling

The research presents a data-driven trading decision system using fuzzy clustering applied to four technical indicators (MACD1, MACD2, S%K, S%D). After data preparation, feature engineering, and standardization, the core methodology employs Fuzzy C-Means (FCM) clustering via the scikit-fuzzy Python library to identify meaningful patterns in market behavior. The clustering algorithm divides the data into two key clusters, corresponding to Buy and Sell market conditions—aligned with the binary nature of trading decisions. The best results were achieved using a fuzziness coefficient of m = 2, ensuring a balance between clear and soft cluster boundaries. The system’s effectiveness was validated using principal component analysis (PCA), which reduced the feature space to two dimensions and visually confirmed well-separated clusters across different currency pairs, reinforcing the model’s robustness.

To translate clustering results into trading actions, the degree of cluster membership for each data point was used to generate Buy or Sell signals, with empirically set thresholds: 0.8 for opening a position and 0.7 for closing. These membership thresholds allowed the system to identify actionable trade signals while filtering out noise (flat market conditions). On testing with the EUR/USD currency pair over a three-year period on daily data, the strategy yielded 40 trades and a net profit of $2,049—amounting to a 20% return on a $10,000 deposit using 0.1 lot per trade. The profit factor reached 2, with a maximum drawdown of only 5.48%, highlighting low risk. When applying modest leverage (1:5), the return would scale to 100% over three years, making the approach attractive even by intraday trading standards.

Performance across other currency pairs confirmed the model’s generalizability, with AUD/USD and USD/JPY delivering even stronger results on the daily timeframe. However, results on the 4-hour timeframe were weaker overall, likely due to increased market noise and the need for model adjustment for intraday dynamics. Importantly, the same entry/exit thresholds were used across all instruments and timeframes, but further analysis showed that optimal thresholds vary by asset. A dedicated experiment mapped the profitability landscape using heatmaps, revealing which combinations of Buy thresholds yielded the best results. For example, EUR/USD performed best when the Buy entry threshold was between 0.90–0.95 and the exit threshold between 0.68–0.73, confirming the potential for further performance tuning.

In summary, the fuzzy clustering model proves to be a powerful and adaptive approach for financial market analysis and trading signal generation. It captures nuanced market behavior, minimizes risk, and delivers strong returns, especially on daily timeframes. The results highlight not only the importance of cluster-based decision logic but also the significant role of parameter calibration in optimizing trading outcomes. With further refinement, particularly for intraday application and asset-specific threshold tuning, this methodology offers a promising foundation for intelligent, data-driven portfolio management.

Conclusions

This study successfully applied the Fuzzy C-Means clustering method to analyse time series, particularly financial instruments. The decision-making system developed using this approach showed good results in identifying trend reversals, which contributed to more efficient buying and selling decisions.

The integration of cluster modelling with traditional technical analysis tools proved to be particularly useful when dealing with complex financial time series. The combination of these methods allowed us to more effectively identify non-linear patterns and process large data sets, which increased the accuracy of trend forecasting.

In particular, testing this approach on the AUD/USD currency pair for 3 years of trading on the daily timeframe demonstrated a profitability of 40%. At the same time, the article shows how to use leverage to further increase profits while remaining within the acceptable risk.

The ability of the models to handle overlapping clusters corresponding to different patterns in price dynamics and to cope with uncertainty about future development of the market makes them highly adaptable to various forecasting tasks in different market conditions. The approach proposed in this study is universal and can be applied to a wide range of financial instruments such as stocks, commodities, futures, etc., and to different timeframes. The importance of using financially-oriented metrics (Net Profit, Max Drawdown, Win Rate, Profit Factor) for assessing the effectiveness of the decision-making system and their advantages over traditional forecasting accuracy metrics are argued.

Although the methodology is self-sufficient for making decisions on financial time series, it is important to optimise the thresholds of membership to clusters for each financial instrument, corresponding to decisions on entering and exiting a position in rising and falling markets. There is also potential in optimising the calculation parameters of technical analysis indicators MACD1, MACD2, S%K, S%D and expanding their number.

First of all, the authors plan to apply filtering of the signals obtained using indicators on higher timeframes, determining permissions for trading directions based on the golden cross method or author’s approach [8], which should eliminate transactions against the dominant trend and improve the trading characteristics of the trading system.

‍

References

1. Tsay, R. S. (2005). Analysis of financial time series (2nd ed.). John Wiley & Sons, Inc,
2. Brockwel, P. J., Davis, R. A. (2016). Introduction to Time Series and Forecasting (3rd ed.). Springer Cham. https://doi.org/10.1007/978-3-319-29854-2
3. Matviychuk, A. (2006). Fuzzy logic approach to identification and forecasting of financial time series using Elliott wave theory. Fuzzy economic review, 11(2), 51-68. https://doi.org/10.25102/fer.2006.02.04
4. Achelis, S. B. (2013). Technical Analysis from A to Z (2nd ed.). McGraw Hill.
5. Elder, A. (1993). Trading for a Living: Psychology, Trading Tactics, Money Management. Wiley.
6. Kuchansky, A., Nikolenko, V., & Rachenko, A. (2015). A method for identifying trends in financial time series based on trend models of forecasting. Management of Development of Complex Systems, 24, 84-89. https://urss.knuba.edu.ua/ua/zbirnyk-24/article-862 [in Ukrainian]
7. Wilson, S. J. (2016). Trend Detection and Pattern Recognition in Financial Time Series [Doctoral dissertation, George Mason University]. MARS. https://doi.org/10.13021/MARS/7317
8. Bakai, Ye. І., Kabachii, V. V., & Maslii, R. V. (2017). Decision-making Model on Financial Time Series Based on a Couple of Moving Averages Using the Assessment of Different Timeframes. Visnyk of Vinnytsia Polytechnical Institute, 1, 70-77. https://visnyk.vntu.edu.ua/index.php/visnyk/article/view/2014 [in Ukrainian]
9. Dash, T., Chitlangia, S., Ahuja, A., & Srinivasan, A. (2021). Incorporating domain knowledge into deep neural networks. arXiv. https://doi.org/10.48550/arXiv.2103.00180
10. Borghesi, A., Baldo, F., & Milano, M. (2020). Improving Deep Learning Models via Constraint-Based Domain Knowledge: A Brief Survey. arXiv. https://doi.org/10.48550/arXiv.2005.10691
11. Kozlovskyi, S., Mazur, H., Vdovenko, N., Shepel, T., & Kozlovskyi, V. (2018). Modeling and forecasting the level of state stimulation of agricultural production in Ukraine based on the theory of fuzzy logic. Montenegrin Journal of Economics, 14(3), 37-53. https://doi.org/10.14254/1800-5845/2018.14-3.3
12. Bielinskyi, A., Soloviev, V., Solovieva, V., & Velykoivanenko, H. (2022). Fuzzy time series forecasting using semantic artificial intelligence tools. Neuro-Fuzzy Modeling Techniques in Economics, 11, 157-198. http://doi.org/10.33111/nfmte.2022.157
13. Nair, B. B., Kumar, P. S., Sakthivel, N. R., & Vipin, U. (2017). Clustering stock price time series data to generate stock trading recommendations: An empirical study. Expert Systems with Applications, 70, 20-36. https://doi.org/10.1016/j.eswa.2016.11.002
14. Kaminskyi, A., Miroshnychenko, I., & Pysanets, K. (2019). Risk and return for cryptocurrencies as alternative investment: Kohonen maps clustering. Neuro-Fuzzy Modeling Techniques in Economics, 8, 175-193. http://doi.org/10.33111/nfmte.2019.175
15. Kononova, K., & Dek, A. (2018). Investigation of traders’ behavioral characteristics: experimental economics methods and machine learning technologies. Neuro-Fuzzy Modeling Techniques in Economics, 7, 148-167. http://doi.org/10.33111/nfmte.2018.148
16. Matviychuk, A., Strelchenko, I., Vashchaiev, S., & Velykoivanenko, H. (2019). Simulation of the Crisis Contagion Process between Countries with Different Levels of Socio-Economic Development. CEUR Workshop Proceedings, 2393, 485-496. http://ceur-ws.org/Vol-2393/paper_423.pdf
17. Enke, D., Grauer, M., & Mehdiyev, N. (2011). Stock market prediction with multiple regression, fuzzy type-2 clustering and neural networks. Procedia Computer Science, 6, 201-206. https://doi.org/10.1016/j.procs.2011.08.038
18. Mehmanpazir, F., & Asadi, S. (2017). Development of an evolutionary fuzzy expert system for estimating future behavior of stock price. Journal of Industrial Engineering International, 13, 29-46. https://doi.org/10.1007/s40092-016-0165-7
19. Shi, Y., Li, B., Du, G., & Dai, W. (2021). Clustering framework based on multi-scale analysis of intraday financial time series. Physica A: Statistical Mechanics and its Applications, 567, Article 125728. https://doi.org/10.1016/j.physa.2020.125728
20. Bandara, K., Bergmeir, C., & Smyl, S. (2020). Forecasting across time series databases using recurrent neural networks on groups of similar series: A clustering approach. Expert systems with applications, 140, Article 112896. https://doi.org/10.1016/j.eswa.2019.112896
21. Long, J., Chen, Z., He, W., Wu, T., & Ren, J. (2020). An integrated framework of deep learning and knowledge graph for prediction of stock price trend: An application in Chinese stock exchange market. Applied Soft Computing, 91, Article 106205. https://doi.org/10.1016/j.asoc.2020.106205
22. Wang, X., Yang, K., & Liu, T. (2021). Stock Price Prediction Based on Morphological Similarity Clustering and Hierarchical Temporal Memory. IEEE Access, 9, 67241-67248. https://doi.org/10.1109/ACCESS.2021.3077004
23. Ozer, F., & Okan Sakar, C. (2022). An automated cryptocurrency trading system based on the detection of unusual price movements with a time-series clustering-based approach. Expert Systems with Applications, 200, Article 117017. https://doi.org/10.1016/j.eswa.2022.117017
24. Sáenz, J. V., Quiroga, F. M., & Bariviera, A. F. (2023). Data vs. information: Using clustering techniques to enhance stock returns forecasting. International Review of Financial Analysis, 88, Article 102657. https://doi.org/10.1016/j.irfa.2023.102657
25. D’Urso, P., De Giovanni, L., & Massari, R. (2021). Trimmed fuzzy clustering of financial time series based on dynamic time warping. Annals of operations research, 299, 1379-1395. https://doi.org/10.1007/s10479-019-03284-1
26. Hartigan, J.A., & Wong, M.A. (1979). A k-means clustering algorithm. Journal of the Royal Statistical Society. Series C, 28(1), 100-108. https://doi.org/10.2307/2346830
27. Dunn, J. C. (1973). A Fuzzy Relative of the ISODATA Process and Its Use in Detecting Compact Well-Separated Clusters. Journal of Cybernetics, 3(3), 32–57. https://doi.org/10.1080/01969727308546046
28. Goddard, J., de los Cobos Silva, S.G., & Gutiérrez Andrade, M.A. (2006). Comparison of k-means and fuzzy c-means using background knowledge. Fuzzy economic review, 11(2), 3-16. https://doi.org/10.25102/fer.2006.02.01
29. Miyamoto, S., Ichihashi, H., & Honda, K. (2008). Algorithms for Fuzzy Clustering. Methods in c-Means Clustering with Applications. Springer. https://doi.org/10.1007/978-3-540-78737-2
30. MetaTrader 4. (2023, April 5). MetaTrader 4 – the most popular Forex trading platform. https://www.metatrader4.com/en
31. Murphy, J.J. (1999). Technical Analysis of the Financial Markets: A Comprehensive Guide to Trading Methods and Applications. ‎New York Institute of Finance.
32. Haryono, A. T., Sarno, R., & Sungkono, K. R. (2023). Transformer-Gated Recurrent Unit Method for Predicting Stock Price Based on News Sentiments and Technical Indicators. IEEE Access, 11, 77132-77146. https://doi.org/10.1109/ACCESS.2023.3298445

‍

‍