380% Hypothetical Return in Perpetuals' BayesShield Backtest — What the Claim Actually Means
The underlying methodology demands a closer look.
What the Engine Does
BayesShield generates automated trading signals by identifying traders who demonstrate consistent performance. The AI layers on top of that signal generation, presumably weighting or filtering inputs from the trader-success model. UpsideOnly itself operates without real capital — it's a simulated environment, which means zero slippage, zero execution latency, and zero liquidity constraints in the tested scenario.
A 380% return in such an environment is a data point, not a result. Hypothetical backtests strip out the primary friction costs that determine whether any strategy survives deployment: order-book depth, partial fills, exchange downtime, and the feedback loop that occurs when signal followers crowd the same entries.
The Baseline Questions
Any backtest claim of this magnitude requires the following parameters for independent evaluation:
- Time period tested. A 380% return over 12 months carries a different Sharpe ratio profile than the same number over 36 months. The announcement does not specify duration.
- Maximum drawdown. Return without drawdown context is a half-metric. Standard deviation of returns, worst peak-to-trough, and recovery time define whether the curve is tradeable or a leverage artifact.
- Over-fitting risk. A model that selects successful traders and then generates signals from their behavior has two degrees of freedom — trader selection and signal synthesis. Each layer increases the parameter space and the probability of curve-fitting to historical noise.
- Walk-forward validation. Was the 380% figure produced on out-of-sample data, or does it reflect in-sample performance? Without this distinction, the number has no predictive weight.
None of these details are present in the available reporting.
Paper Returns vs. Execution Reality
UpsideOnly's "risk-free" framing confirms the environment carries no capital at risk. That is a controlled testing ground — useful for signal iteration, but structurally disconnected from live market conditions. The gap between a paper backtest and deployed capital is not linear; it's governed by the interaction of position sizing, liquidity consumption, and market microstructure.
Traders evaluating AI signal providers should treat the 380% figure as a starting hypothesis requiring independent validation — not as evidence of edge. The relevant metric is not return, but risk-adjusted return after execution costs, measured over a statistically significant sample with out-of-sample confirmation.
Perpetuals has made a claim. The supporting data to make that claim actionable has not yet been published.