Stock return forecasting is of utmost importance in the business world. This has been a major topic of research for many academicians for decades. Recently, regularization techniques have reported significant increase in the forecast accuracy of the simple regression model. Still, more robust techniques are desired which can further help improve stock prices predictions. Furthermore, it is important to recommend top stocks rather than predicting exact stock returns. The technique should be scalable to very large datasets. The present paper proposes a normalization technique which results in a form of regression that is more suitable for ranking the stocks by expected returns. The ranking is done out of the comparison between the stocks in the previous quarter over a big set of important fundamental, technical and general indicators. Two large datasets consisting of altogether 946 unique companies listed at Indian exchanges were used for experimentation. Stochastic Gradient Descent technique is used in this work to train the parameters, which allows scalability to even larger datasets. Five different metrics were used for evaluating the different models. Results were also analysed subjectively through plots. The returns obtained were higher than other popular models and benchmark indices.

股票回报预测在商业世界中至关重要。几十年来，这一直是许多院士的主要研究课题。最近，正则化技术报告了简单回归模型的预测准确性显着提高。尽管如此，仍需要更强大的技术，以进一步帮助改进股票价格预测。此外，重要的是推荐顶级股票而不是预测确切的股票回报。该技术应该可以扩展到非常大的数据集。本文提出了一种归一化技术，该技术产生一种更适合按预期收益对股票进行排名的回归形式。该排名是根据上一季度股票在一系列重要的基本面、技术面和一般指标上的比较得出的。由在印度交易所上市的总共 946 家独特公司组成的两个大型数据集用于实验。在这项工作中使用随机梯度下降技术来训练参数，这允许扩展到更大的数据集。五个不同的指标用于评估不同的模型。还通过绘图对结果进行了主观分析。获得的回报高于其他流行模型和基准指数。