Introducing an Innovative Approach to Mitigate Investment Risk in Financial Markets: A Case Study of Nikkei 225

The study addresses the challenge of accurately forecasting stock market trends, a task complicated by the noisy, nonlinear, and chaotic nature of financial time series and the influence of numerous factors (e.g., liquidity, investor behavior, and news). Traditional time-series approaches rely on static patterns and struggle with market volatility and the multitude of influencing variables. Advances in machine learning, particularly support vector regression (SVR), offer improved generalization, robustness to outliers, and effectiveness with limited samples. The research question is whether a hybrid SVR model optimized with modern meta-heuristics can deliver more accurate stock price forecasts, thereby mitigating investment risk. The purpose is to develop and evaluate a hybrid forecasting framework for the Nikkei 225 index using 2013–2022 data, integrating multiple optimizers—moth flame optimization (MFO), artificial bee colony (ABC), and genetic algorithm (GA)—and to compare their effectiveness. The study emphasizes the importance of robust prediction for informed decision-making and economic stability, while highlighting the need for optimization techniques that avoid local minima and balance exploration and exploitation. It also identifies gaps in understanding dataset characteristics affecting algorithm performance, generalizability across conditions and geographies, integration of external variables, and the need for standardized evaluation metrics. The contributions include proposing an MFO-optimized SVR (MFO-SVR) and benchmarking it against GA-SVR, ABC-SVR, baseline SVR, MLP, and ARIMA, with standardized metrics (MAE, MAPE, MSE, RMSE).

Recent work extensively explores ML for stock market forecasting. Agrawal proposed deep learning-based nonlinear regression outperforming traditional ML on Tesla and NYSE data. Petchiappan et al. applied logistic/linear regression to media and entertainment stocks, demonstrating practical gains. Sathyabama et al. emphasized external variables (e.g., news) and proposed a Naïve Bayes-based approach. Menaka et al. surveyed ML algorithms (ensembles, SVM, RF, boosted trees) for trend prediction across exchanges. Demirel et al. compared MLP, SVM, and LSTM on Istanbul Stock Exchange data for opening/closing price prediction, while Tembhurney et al. benchmarked SVM and RF on Nifty 50 forecasting. The literature underscores ML’s utility but reveals gaps: lack of cohesive comparative frameworks across varied datasets/conditions, limited integration of exogenous variables (geopolitics, macro indicators, sentiment), sparse discussion of real-world deployment impacts, and insufficient handling of obsolete/inconsistent data. This study responds by proposing a hybrid SVR optimized via MFO, ABC, and GA, trained/tested on Nikkei 225 (2013–2022), and evaluated with standardized metrics to provide rigorous, comparable evidence of efficacy.

Dataset: Daily Nikkei 225 OHLC and volume data from 2013 to 2022 were collected. Rigorous cleaning addressed missing/inconsistent entries and outliers through inspection and preprocessing. Features were min–max normalized (typically to [0,1]) to stabilize training. The dataset was split 80%/20% into train/test. Candlestick representations (OHLC) informed descriptive analysis. Bias considerations: The authors note potential biases stemming from data completeness/accuracy, feature selection, and subjectivity in interpreting candlestick patterns. They applied validation and sensitivity analyses and continuous monitoring to mitigate such biases. Model: Support Vector Regression (SVR) was used as the core predictive model due to its robustness and generalization. The standard ε-insensitive loss formulation with regularization (C), slack variables, and kernel functions was adopted. The decision function employs kernel K(·,·) with tuned hyperparameters. Optimizers for SVR hyperparameters: - Genetic Algorithm (GA): Population-based search using selection, crossover, and mutation to optimize SVR hyperparameters. While heuristic and not guaranteed to find global optima, GA can yield good solutions at reasonable cost; it may be computationally heavy for large datasets. GA-selected SVR hyperparameters: kernel linear; gamma 0.5; C 10; epsilon 0.05. - Artificial Bee Colony (ABC): Swarm-based approach inspired by honeybee foraging, iteratively improving candidate solutions via employed, onlooker, and scout phases. ABC-selected SVR hyperparameters: kernel linear; gamma 0.1; C 20; epsilon 0.05. - Moth Flame Optimization (MFO): Swarm-based optimizer modeling moth spiral flight around flames, designed to balance exploration/exploitation and avoid local optima with adaptive reduction in flames over iterations. MFO-selected SVR hyperparameters: kernel linear; gamma 0.5; C 10; epsilon 0.1. Baselines: Autoregressive Integrated Moving Average (ARIMA) and Multilayer Perceptron (MLP) were implemented for comparison alongside vanilla SVR. Evaluation metrics: MAE, MAPE, MSE, and RMSE were used for quantitative assessment on both train and test sets. Statistical profiling: The OHLCV series (n=2442) were summarized (means around 20,813 for open/close with std ≈4,764–4,777; volume mean ≈3,730, std ≈1,985; skewness near 0.17 for prices and 1.91 for volume).

- On the test set, MFO-SVR achieved the best performance among all models: RMSE 230.60, MAPE 0.70, MAE 197.53, MSE 53,175.49. - Comparative test-set results: - ARIMA: RMSE 348.19; MAPE 1.09; MAE 307.16; MSE 121,236.11. - MLP: RMSE 314.20; MAPE 1.02; MAE 287.83; MSE 98,721.97. - SVR: RMSE 291.21; MAPE 0.88; MAE 247.12; MSE 84,803.10. - GA-SVR: RMSE 275.18; MAPE 0.79; MAE 220.27; MSE 75,724.66. - ABC-SVR: RMSE 259.20; MAPE 0.75; MAE 211.43; MSE 67,182.42. - MFO-SVR: RMSE 230.60; MAPE 0.70; MAE 197.53; MSE 53,175.49. - Training-set trend mirrored test results, with MFO-SVR outperforming: RMSE 97.55; MAPE 0.38; MAE 68.98; MSE 9,516.14. For comparison: ARIMA (RMSE 255.34; MAPE 1.01; MAE 184.44; MSE 65,199.22), MLP (RMSE 226.20; MAPE 0.91; MAE 168.52; MSE 51,164.83), SVR (RMSE 185.35; MAPE 0.75; MAE 136.94; MSE 34,352.80), GA-SVR (RMSE 142.10; MAPE 0.56; MAE 103.95; MSE 20,193.55), ABC-SVR (RMSE 124.29; MAPE 0.47; MAE 85.23; MSE 15,448.45). - The standardized evaluation demonstrated consistent gains from optimizer integration into SVR, with MFO providing the largest improvement in both error magnitude and percentage terms. - The reported MAPE of 0.70% on the test set evidences high-accuracy forecasting for Nikkei 225 closing prices over the 2013–2022 period.

The findings confirm that optimizing SVR via meta-heuristics significantly improves predictive accuracy over baseline SVR and traditional models (ARIMA, MLP). Among the hybrid approaches, MFO-SVR delivered the lowest errors across RMSE, MAE, MAPE, and MSE, indicating better generalization and effective avoidance of local minima through MFO’s exploration–exploitation balance. Practically, the enhanced accuracy supports applications in investment decision support, algorithmic trading, and risk management, where timely and precise forecasts can reduce volatility exposure and inform portfolio adjustments. The study demonstrates that, with rigorous preprocessing and standardized metrics, a hybrid SVR can model nonlinear market dynamics more effectively than conventional techniques, thereby addressing the research objective of improving stock price forecasts to mitigate investment risk. The results also reinforce the importance of hyperparameter tuning strategies and swarm intelligence methods in financial time-series modeling. However, the authors acknowledge the dependence on historical data and potential limitations in capturing exogenous shocks, indicating that integrating external information (news sentiment, macro indicators) could further fortify robustness and real-world applicability.

The paper introduced a hybrid stock price forecasting framework centered on SVR optimized by MFO and benchmarked against ABC-SVR, GA-SVR, baseline SVR, MLP, and ARIMA using Nikkei 225 data (2013–2022; 80/20 train–test split). MFO-SVR achieved the best performance, notably a test MAPE of 0.70% and the lowest RMSE/MAE/MSE among competitors, demonstrating high accuracy and practical relevance for investment decision-making. The study contributes standardized evaluation and a comparative analysis of optimization strategies for SVR in financial forecasting. Future work includes: expanding to additional markets for generalizability; integrating real-time data and exogenous variables (macroeconomics, sentiment); exploring alternative or ensemble optimizers; and advancing real-time prediction capabilities, all to enhance robustness, interpretability, and actionable deployment in dynamic financial environments.

- Dependence on historical data may limit responsiveness to sudden regime shifts or unprecedented events, potentially degrading accuracy during crises or structural breaks. - Potential biases in data quality/completeness and feature selection; candlestick-based pattern interpretations can introduce subjectivity. - Generalizability across asset classes, market regimes, and geographies is not established; results are specific to Nikkei 225 over 2013–2022. - Transaction costs, liquidity constraints, and market impact are not modeled, so live trading performance may differ from backtests. - Model interpretability is limited relative to simpler econometric approaches; further work is needed to enhance transparency and stakeholder trust.