High accuracy intraday stock candle prediction model

- Output shape: 5 floats per candle window, z-score normalized per rolling window (not global) - Fields: highest_high_dist (-3, +3), lowest_low_dist (-3, +3), directional_move (-1, +1), realized_range (0, +3), prediction_confidence (0, 1) - Average holdout directional accuracy (all datasets) ≥ 58% (above random baseline, in holdout set) - Reproducible: same input data + same seed = same output, every time - Must provide import/load functions that return a numpy array or tensor in the exact shape above - No future leakage: all features computed from past data only, validated by the live parity test - Minimum holdout directional accuracy (each dataset) ≥ 53% - Average holdout directional correlation or Information Coefficient between predicted directional_move and actual future directional move must be at least plus 0.12 - Minimum per-dataset directional correlation or Information Coefficient must be at least plus 0.06 - Spearman Rank IC must be positive on the majority of walk-forward holdout folds, so the predictions are not only directionally accurate but also meaningfully ranked - For predictions where the direction is correct, mean percentage magnitude error must be 0.35 percent or lower. For example, if the real move is down 1 percent and the model predicts down 0.5 percent, that counts as 0.5 percent error - Prediction confidence must be meaningful: the highest-confidence prediction group should perform better than the full holdout average, preferably by at least 3 percentage points in directional accuracy - Reproducible: same input data plus same seed plus same configuration must produce the same output every time - Must provide import and load functions that return a NumPy array or tensor in the exact shape above - No future leakage: all features must be computed from past data only, validated by a live parity test - Chronological walk-forward method must be used, with train, test, validation, and holdout splits. No random time-series shuffling is allowed - I will provide the data, including assets, timeframes, dates, and final split logic - The developer can use any suitable method, including LSTM, BiLSTM, GRU, Transformer, Temporal CNN, XGBoost, LightGBM, CatBoost, hybrid models, ensemble models, or any other serious method - There is no need to build the whole project from zero. The developer can choose a strong paper, existing repository, or proven architecture and adapt it to this task, as long as the final result satisfies the output shape, walk-forward validation, no-leakage, reproducibility, and holdout performance requirements - This is not a full trading bot. No exchange integration, order execution, portfolio management, or trading UI is required. The goal is only to build a reliable candle-prediction feature layer that will feed another decision layer Walkforward will be used for development and test stage. Train/test is classical 70/30 split that will be used for development, validation will be used a few times as main test to detect issues and improvement patterns. Holdout will be used for final testing stage and will be ran only 1 or 2 times at live after developer thinks that code is finally ready. Model will be trained from 0 and together and those will be final results. Finally live test will be done for a few days to %100 confirm that there is no future leakage, there isn't any accuracy condition at live, it only has to be same with backtest results (same range will be downloaded and tested and compared)