Educational content from the Criptospace desk. Not investment advice, and not a solicitation to invest.
Most retail conversations about trading start and end with the signal: which model predicts direction best, which indicator catches the turn first, which strategy has the highest win rate. Inside a systematic desk, the signal is treated differently — as one input among several, and rarely the one that determines whether a strategy survives.
Why signal accuracy is the wrong scoreboard
A signal that is right 55% of the time can still lose money if position sizing is wrong, execution costs eat the edge, or the strategy is fully exposed exactly when correlations break down. Conversely, a modest edge survives and compounds when it's paired with disciplined sizing and clean execution. This isn't a controversial idea in quantitative finance — but it's consistently underweighted in how crypto trading gets discussed publicly, where the conversation still centers on calls and predictions rather than the infrastructure that turns an edge into a return.
The desk treats every strategy as three separable components: the signal, the risk model that sizes it, and the execution layer that expresses it in the market. Each is engineered and tested independently, because each fails independently.
Regimes change faster than models retrain
Crypto markets rotate between regimes — trending, mean-reverting, thin and fragmented, deep and correlated — often within the same week. A model tuned on the last regime doesn't fail loudly; it fails quietly, bleeding through a string of small losses that look like normal variance until they aren't.
Rather than trying to predict the next regime, the desk's risk layer is built to detect the current one from realized volatility, dispersion and liquidity conditions, and to adjust exposure accordingly:
- Volatility targeting. Position size scales inversely with realized volatility, so a strategy's risk contribution stays roughly constant even as the underlying asset's behavior changes.
- Correlation-aware sizing. Exposures across strategies and assets are netted against a live correlation matrix, not treated as independent — the biggest drawdowns in crypto come from positions that looked diversified on paper and moved together in practice.
- Drawdown throttles. Every strategy carries a mechanical de-risking curve: as realized drawdown increases, allocated capital steps down automatically, before a human has to intervene.
None of this predicts the market. It just makes sure the desk is never fully exposed to a regime it hasn't recognized yet.
Execution is part of the strategy, not an afterthought
A well-sized position still loses money if it's poorly executed. Crypto liquidity is fragmented across dozens of venues, order books are thin outside the top few pairs, and the venue with the best price a minute ago may not be the venue with the best price now. Execution decisions — which venues to route to, how to slice an order, when to hold back — are as consequential as the signal that triggered the trade.
This is why the desk's execution layer is built, not bought: order routing is idempotent by design, so a dropped connection or an exchange outage can't result in a duplicate fill or a silent miss. Orders are sliced against real-time depth rather than a fixed schedule, and every fill is measured against arrival price to keep slippage visible — tracked trade by trade, not estimated after the fact.
Where AI actually helps — and where it doesn't
"AI-coordinated execution" gets used loosely across the industry, so it's worth being precise about what that means on this desk. Machine-learned models are used for the narrow, well-bounded problems they're good at: classifying the current volatility regime, estimating short-horizon liquidity, and optimizing how an order should be sliced across venues and time. They are not used to generate directional predictions that override the risk model, and they don't get to increase exposure on their own — sizing limits and drawdown throttles sit above the model, not below it.
That separation matters. A model that's occasionally wrong about direction is a normal cost of doing business. A model that's occasionally wrong about how much capital to risk is a different category of problem — so that decision stays governed by fixed, auditable rules rather than a prediction.
What this looks like in practice
Put together, the framework is less about finding better signals and more about making sure a good-enough signal survives contact with real markets:
- Detect the current regime from realized data, not a forecast.
- Size positions to a constant risk budget, adjusted for correlation.
- Throttle capital mechanically as drawdown increases.
- Execute against live depth with routing that's resilient to venue failure.
- Keep the model's authority bounded to what it's actually good at.
It's a less exciting story than a winning signal, but it's the difference between an edge that shows up in a backtest and one that survives a live market that doesn't know what your model expects it to do next.