Every LSTM in this paper loses on MAPE to forecasting tomorrow's close at today's price. That result covers three Indonesian stocks over 178 days, and the authors acknowledge it. For a practitioner, the question is how much evidence their feature rankings can offer.
What Zahra and Zaky built
Forecast error is the payoff being tested here. There is no trading strategy. The LSTM forecasts next-day return, which is then converted into tomorrow's price; all scores are calculated in price units. Pruning has a reasonable premise: technical indicators often reuse the same closing price, so removing redundant inputs should reduce computational cost and overfitting.
Zahra and Zaky examine ADRO, ASII and TLKM, respectively an energy, an automotive and a telecom name on the IDX. Their adjusted daily OHLCV data and Jakarta Composite Index (JCI) series come from yfinance, covering May 2021 to May 2026. Each stock has 1,179 usable rows. The starting set contains twenty features. Almost all are expressed relative to that day's close or a 20-day volume average, keeping test-period values within the training range. This conversion reduced the proportion of normalised test values outside [0, 1] from 15.11% to 60.39% down to 0.70% to 3.57%. Next-day return remains the target, with the price reconstructed afterward.
The selector uses two scoring rules from the same model. A 500-tree Random Forest predicts next-day return, matching the LSTM's target, and is fitted on the 825 training days using three seeds. Each fit yields impurity importance, called RF Importance in the paper, and mean absolute SHAP, a score of each feature's contribution to the model's predictions. The authors average each score across seeds, then normalise it as a share of the total. Features survive at or above 1/20, or 5%. They then train one fixed LSTM on all 20 features, the RF selection and the SHAP selection, using five seeds for each. Evaluation uses the August 2025 to May 2026 test block.
For ADRO, ASII and TLKM, RF retains 12, 11 and 10 features respectively. SHAP retains 9 in each case. Across every scheme, test MAPE ranges from 1.76% to 1.90%. The authors describe the main benefit as "a simpler model input, not better prediction accuracy." I agree with the latter part. Their interpretation of the rankings asks more of the evidence than it can support.
Does anything beat the last close?
Nothing does on MAPE. The naive forecast records 1.72%, 1.85% and 1.87% for the three stocks; every scheme finishes 0.01 to 0.11 points behind. ASII supplies the only edge over naive, measured by RMSE: 156.77 to 157.67 versus 157.74. Even there, naive has the lower MAE, 110.26 versus 110.70 to 111.48. The largest improvement, 0.97, amounts to less than one rupiah for a stock trading between 3,151 and 7,044.
All nine stock and scheme pairs have directional accuracy below 50%, ranging from 41.80% to 49.55%.
The table omits directional accuracy for the naive forecast. We also did not find an explanation of how the sign-match rule handles an actual return of zero. Counting zero-return days as misses would make 50% the wrong benchmark. The authors interpret the one-day lag in their price plots as evidence that the LSTM forecasts returns close to zero. These figures fit that reading. We encountered the same tie with a zero-return random walk in next-day Straits Times graph forecasts.
Half the inputs, about 4% fewer weights
The input count shrinks by 40% to 55%. The parameter count moves from 125,761 to between 120,129 and 121,665, a reduction of 3.3% to 4.5% that the authors report plainly. Those counts also reveal something an implementer needs to know. Eleven additional inputs contribute 5,632 parameters, or 512 each. Four gates multiplied by 128 units gives 512, implying a first LSTM layer of width 128. The 12-feature model's 121,665 parameters agree with that calculation. Our own arithmetic (125,761 minus 20 times 512) leaves 115,521 weights independent of the input count. The inputs are simpler. Computational savings were never measured.
A selector with negative out-of-sample R²
About 0.86 is the Random Forest's training R². Its out-of-bag R² ranges from −0.054 to −0.008 across all three stocks. The authors disclose both and invoke the latter to explain why no selected subset improves the forecasts. That explanation holds up.
The investor implication they draw later is harder to defend against the same result. Keeping volume and JCI features, they argue, suggests that trading activity and market conditions "are more useful inputs than trend indicators calculated from the same closing price." Yet the authors also explain the eliminations through redundancy and credit-splitting. For discarded features, the average maximum correlation with another feature is 0.91 to 0.97; for retained features, it is 0.69 to 0.74. As they acknowledge, Random Forest scores divide credit among near-duplicates. Both selection methods remove SMA 20 and BB Middle, identical by construction. A fixed 5% threshold makes features with close twins more likely to be dropped.
I object to the further inference. The paper says volume and the JCI "carry information that is not contained in the stock's own price" and "therefore have a higher correlation with returns." Retained features have average absolute correlations with return of 0.037 to 0.098. For discarded features, the range is 0.018 to 0.052. There is overlap: 0.037 is below 0.052. Both ranges are tiny beside an out-of-bag R² of −0.054 to −0.008. My reading is that volume and the JCI largely survive because the 20-feature set contains nothing resembling them. Their survival reflects the construction of that set.
I would change my view if the same five survivors (Open, Volume, JCI High, JCI Low, JCI Volume) ranked first under permutation importance on the validation block, with a positive fit there.
Where a rerun still requires judgment
The authors specify most of the procedure precisely, including seeds, patience and learning-rate halving. Some choices remain unresolved:
- "Relative difference from the closing price" comes without a formula, leaving us to choose the denominator and sign.
- RSI uses average gains and losses over 14 days, without specifying simple or Wilder smoothing.
- The treatment of zero-return days in directional accuracy remains unclear, as discussed above.
- The authors say the data "can be retrieved again using the same parameters," although yfinance adjusts history again after subsequent dividends. A rerun therefore drifts from the original series.
The comparison criterion is thin as well. To win, a scheme's MAPE improvement must exceed the inter-seed standard deviation, 0.01 to 0.02 points across five seeds. Significance testing is left to future work. Every comparison also depends on a single chronological split.
We could not rerun the study. Our data covers US equities, includes neither IDX listings nor JCI volume, and ends before the paper's sample, which extends into 2026. Substituting US stocks would produce a different experiment.
SHAP's 9 inputs keep pace with all 20 on ASII and TLKM
The 9-feature SHAP selection finishes within about 0.01 points of all 20 inputs for ASII and TLKM: 1.87% versus 1.86%, and 1.88% versus 1.89%. ADRO behaves differently. Its full-input MAPE is 1.76%, compared with 1.83% for RF and 1.81% for SHAP. Those gaps exceed the seed spread. ADRO also spans the widest price range, with a price standard deviation of 36.5% of its mean. Across every scheme, its validation loss is 134% to 136% above training loss.
Without a strategy or a cost model, these results say nothing about profitability. They do show that SHAP removes more inputs than impurity importance (9 features versus 10 or 11) at a MAPE cost no greater than 0.01 points on ASII and TLKM. The selector behind that result cannot forecast the target whose features it ranks.