Wavelet-Based Short-Horizon Forecasting of BTCUSDT Returns

Overview

Implementation

Toolboxes

Deep Learning Toolbox for the neural stages, Wavelet Toolbox for the transforms, Parallel Computing Toolbox optionally for feature extraction and multi-seed training

Windows

A 256-bar lookback with forecast horizons of 30, 60 and 128 one-minute steps

Splits

Chronological 70/15/15 with a configurable buffer between blocks

Seeds

Three repeated seeds per model, aggregated into the test summary

Wavelet settings

Dual-tree preferred with a nearsym5_7 level-one filter; a four-level db4 decomposition as fallback; identity as the last resort, and a switch that forbids falling back at all

Models

Zero, last-value, mean and ridge-style multistep linear baselines, plus raw, wavelet and dual-branch residual temporal-convolutional networks

Metrics

MAE, RMSE, stepwise directional accuracy, cumulative directional accuracy and Pearson correlation, all in the return domain

Components

config/default_config.m, experiment_presets.m

The single configuration struct and the named presets, including the smoke preset

scripts/stage01–stage03

Loading and cleaning one-minute bars, computing log returns, and building chronological splits

scripts/stage04–stage05

Overlapping lookback and forecast windows, and the wavelet feature matrices

scripts/stage06–stage09

Baseline fitting and the raw, wavelet and dual-branch neural stages

scripts/stage10–stage12

Evaluation, figures, and the validation-based hyperparameter sweep

src/features/

The wavelet adapter with its preferred, fallback and identity paths

src/models/

The residual regressor trainer plus the custom forecast-regression and feature-slicing layers

src/normalization/

Normalizers fitted on the training split and applied unchanged downstream

src/evaluation/

Return-domain scoring and the across-seed aggregation

src/plotting/

Report styling, representative-example selection, figure export, and the figure source-data export

src/utils/create_run_context.m, create_report_bundle.m

Per-run output context and the report bundle assembled from it

src/io/download_binance_klines_csv.m

Optional direct download of the one-minute bar history

Stages

1. stage01–02 Load and clean the one-minute CSV, deduplicate, sort, report missing minutes, and compute log returns.
2. stage03 Build chronological train, validation and test splits with the configured buffer.
3. stage04 Generate overlapping lookback and forecast windows for each horizon.
4. stage05 Extract wavelet features, recording which transform path was used.
5. stage06 Fit the four baselines.
6. stage07–09 Train the raw, wavelet and dual-branch residual networks across seeds.
7. stage10 Evaluate every model in the return domain and aggregate across seeds.
8. stage11–12 Render figures and, optionally, run the validation-based hyperparameter sweep.

Reproducibility and validation

• A deterministic fixture dataset drives an end-to-end smoke run, so the twelve stages can be exercised without the real bar history.
• The wavelet adapter records the transform path it took, and require_preferred_method makes a fallback an error rather than a silent substitution.
• All scoring happens in the original return domain, so a normalization choice cannot flatter one model relative to another.
• Each run writes its own directory with a configuration snapshot, metrics tables, figures, model artifacts, metadata and the repository commit hash.
• Figure source data is exported alongside every figure — the CSV exports behind the published figures come from this step, and are the data available on the research page.
• 18 test files cover the stages individually plus end-to-end integration and the download wrapper.

Availability