Financial market prediction built from scratch. A proprietary architecture trained on raw OHLCV data across 16 cryptocurrency pairs, with live paper trading running continuously since March 2026.
Most financial ML treats price data as a flat time series. We treat it as a multi-scale signal — decomposing raw candle data using the same wavelet-based tokenization that powers our EEG seizure prediction work. The core insight: financial markets and brainwaves share a common structure — information encoded across multiple temporal scales simultaneously.
The architecture goes beyond standard wavelet decomposition. Multiple proprietary enhancements to both the tokenization and the sequence model produce measurable gains at each generation — validated through rigorous backtesting and live paper trading.
Six generations of iterative refinement. Each generation builds on the last, with architectural changes driven by empirical results rather than theory alone.
Wavelet tokenizer + base sequence model. Established the core pipeline from raw OHLCV to directional predictions.
75.5% direction accuracy
Refined tokenizer vocabulary and model capacity. Incremental gains confirming the approach.
75.6% direction accuracy
Extended training and architecture tuning. Diminishing returns on the base approach pointed toward structural changes.
75.7% direction accuracy
Proprietary structural innovation. A qualitative change in how the model processes temporal information. Live paper trading since March 2026.
76.0% direction accuracy · 88.2% weighted trading accuracy
Tested further refinements within the Gen 4 framework. Revealed a critical insight: optimising the training objective harder can degrade the actual goal. This finding shaped our research direction going forward.
Parked — insight carried into Gen 6 design
Extended architecture search. Integration of a new structural primitive broke through the Gen 4 plateau on held-out evaluation. Proprietary; details not published while live integration is validated.
68.8% direction accuracy on held-out cross-candle eval · live integration in progress
Lower training loss does not equal better real-world performance. We confirmed this across three independent experiments — each improved the proxy metric while degrading the actual prediction target. This finding has direct implications for how financial ML models should be evaluated and selected.
The Gen 4 model runs live paper trades on 16 cryptocurrency pairs (4-hour candles via Binance). All trades are logged, timestamped, and verifiable. No cherry-picking, no hindsight adjustments.
Why show a sub-50% win rate? Because honest science matters more than impressive claims. The model's 76% direction accuracy on historical data hasn't yet translated to equivalent live performance — the gap between backtesting and live trading is real, and understanding it is part of the research.
What the paper trading does demonstrate: the model identifies which assets move most (3.8x better than random selection across 16 pairs), even when directional calls are noisy. This is a building block, not a finished product.
The same wavelet tokenization architecture powers both FinForm (financial prediction) and NeuroWave (EEG seizure prediction). Different input domains, shared core insight: temporal signals with multi-scale structure benefit from wavelet decomposition before sequence modelling.
This isn't a coincidence — it's the thesis. We believe this approach generalises to any domain where information is encoded across multiple timescales: seismology, audio processing, industrial sensor monitoring, climate modelling, and more.
All models trained and deployed on local hardware. No cloud compute. No third-party APIs.
FinForm is proprietary research. For collaboration or enquiries:
Contact — finform@hyperreal.com.au