Models that learn from noisy data
ANC Labs is an independent research lab working on one question: how do you make machine learning hold up when the signal in financial data is barely distinguishable from noise?
- higher signal efficiency than a standard xLSTM (synthetic tests, 5 cells)
- 16–36%
- measurements in the synthetic lab
- 1,583
- symbols in the universe, delisted coins included
- 832
Why financial data is hard
Only a sliver of tomorrow's return on a coin is predictable. The rest is noise. Tails are fat, assets move together, and the market changes character from one period to the next.
Flexible models pick up the wrong thing in this setting: there is just enough data to memorise the noise instead of the signal. In our lab, a standard xLSTM reached its best validation score within the first two epochs in 86% of runs. What came after was memorisation.
Research in detail- DA-xLSTM
- standard xLSTM
- linear ceiling
A market where we know the signal
On real data you can't tell which model is more robust to noise, because nobody knows the true signal. So we first built a data generator calibrated to the statistics of the real market: fat tails, cross-sectional correlation, GARCH-style volatility, regime shifts and a funding component.
We set the strength of the signal ourselves, which means we can measure directly how much of it a model recovers.
13 of 13 calibration targets within tolerance.
How the lab worksA noise-robust xLSTM
xLSTM's exponential gating and large memory are strengths on clean data. At low signal-to-noise, a single outlier can dominate that memory for many steps, and spare capacity gets spent on noise. DA-xLSTM goes after this with three changes: a clipped training target, a deliberately smaller hidden state, and an average over five models.
We tested each change on its own first, against a hypothesis and pass criterion written down before the run. The combined model beat the standard xLSTM in all 5 test cells.
Architecture and ablationmilliways01
The production system at the centre of the research programme: a market-neutral, cross-sectional portfolio built each day on roughly the hundred most liquid Binance perpetuals. It combines two families of signals, one from order flow and one from funding, and is in a live forward test on an engine that runs on real-time market data and order books.
- 01Market data
- 02Recorder
- 03Decision engine
- 04Data validation
- 05Monitoring
- Sharpe ratio
- 1.76
- annual compound return
- 25.3%
- maximum drawdown
- −14.4%
How we test
It's easy to fool yourself with noisy data, so every result goes through the same rules.
- 01
No look-ahead
Every feature is known at decision time. Decide on day t, trade on day t+1.
- 02
After costs
Nothing counts until fees, measured spreads and market impact are taken out.
- 03
No survivorship bias
The universe is rebuilt every day from that day's information, including coins that were later delisted.
- 04
Pre-registration
Hypothesis, metric and pass threshold are written down before a test runs.
- 05
Independent reproduction
Every number that drives a decision is produced a second time, independently.
- 06
Mutation testing
We check that critical safeguards really catch errors by breaking the code on purpose.
155 experiments, 10 families
Every idea we've tried is on record: architectures, features, cross-sectional signals, risk management, data sources, execution. Each dot is one experiment. The 1,583 model runs from the synthetic lab come on top of these.
Built for the market
ANC Labs is a one-person, independent lab, and the research is a means, not the end. The goal is to build systems that earn sustainable returns in live markets, after costs.
So every idea faces the same question: does it hold up in a live market, with real costs? Every layer, from model development and data pipelines to test infrastructure and the live engine, is built around that goal. milliways01 is the first production system on that path.
The lab and its founder- 01
Returns after costs
An idea only moves forward if it still pays after fees, spreads and market impact.
- 02
Proven live
No system moves to capital before a forward test on real-time market data.
- 03
Risk first
Drawdown, capacity and cost sensitivity are measured before capital grows.