Why wavelets
A dual-tree complex wavelet history exposes structure across scale and time at once, giving the network a richer, image-like input than a single forecast sequence.
Research series
Wavelet-GAN Bitcoin Forecasting
A multi-year line of work asking one question: can wavelet-conditioned generative models extrapolate short-horizon Bitcoin price changes well enough to matter?
Four linked reports trace the arc from the original concept, through the first WaveGAN experiment and a refined architecture, to a controlled re-evaluation that holds the early optimism to account. Each part below opens its full report.
The theme
One question, followed across four reports.
Bitcoin's price behaves like a random walk, but its minute-to-minute changes hold a stubbornly regular distribution. This series asks whether a wavelet representation, paired with a conditional GAN, can turn that regularity into a useful short-horizon forecast, and tracks the answer honestly as the work matured.
Why a GAN
The target is a plausible continuation of a stochastic signal, not an exact path. A conditional adversarial objective is well suited to learning that distribution rather than a single point estimate.
How to read it
The first two reports build and test the idea; the third documents the architecture in full; the fourth re-evaluates it against simple baselines and reports exactly where it stands.
The arc
From concept to controlled evaluation.
Four reports in sequence. Each entry opens the complete original page, with its prose, figures, tables, and references preserved.
-
Trading AI Proposals
The founding concept paper. It argues that while Bitcoin's price looks like a random walk, its minute-to-minute price changes hold a near-invariant distribution a network can exploit, and proposes pairing wavelets with a GAN to extrapolate it.
-
WaveGAN BTCUSDT Extrapolation Report
The first concrete build: a MATLAB wavelet-augmented conditional GAN evaluated in the wavelet domain across 30-, 60-, and 128-minute horizons, beating an identity baseline by up to 50.76% at 30 minutes.
-
Wavelet GAN Trading Prototype
The architecture deep-dive: a full write-up of the wavelet-conditioned generator and discriminator, the compound loss, and the training and evaluation design behind the prototype.
-
Wavelet Feature Forecasting Study
A controlled re-test that holds the earlier optimism to account: against matched baselines, ridge regression, and residual TCNs, the neural wavelet variants do not win — a clean, documented answer that reframes where to look next.