Building a Stock Direction Classifier with CNN and LSTM
Summary
This legacy platform tutorial outlines a workflow for predicting stock direction with a hybrid convolutional and recurrent neural network. It labels each observation according to whether the following ten-day return is positive or negative, builds rolling input sequences from selected features, and describes a CNN layer followed by an LSTM, dropout, and dense output layers. The example uses windows of 50 rows and five features, and the final output is interpreted as an upward-move probability.
The proposed rule buys or holds when that probability exceeds 0.5 and sells or stays out when it falls below 0.5. The tutorial also describes training, validation-set prediction, aligning outputs with dates, and a simulated backtest with fees and slippage. It presents a construction recipe, not evidence of predictive performance: no results, benchmark comparison, or robustness analysis are supplied. The page warns that its platform modules and resources are outdated, and leaves feature choice, model settings, and validation quality to the implementer.
Key ideas
- The tutorial labels observations by the sign of the subsequent ten-day stock return.
- It uses fixed-length feature sequences as input to a CNN-LSTM model.
- The example maps five features across 50 rows to an estimated probability of an upward move.
- Its trading rule enters or holds above a 0.5 probability threshold and exits below it.
- It describes a backtest workflow but provides no performance evidence or robustness checks.
- The platform instructions are explicitly identified as outdated.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.