Most backtests are dishonest. Not intentionally — the industry just normalised a set of practices that make historical performance look better than it ever was. Signal computed on bar t trades on bar t. Parameters optimised over the full dataset, including data the strategy would never have seen live. Sharpe ratios that no live strategy has ever reliably replicated.
The academic tail
We had been reading the recent wave of AI trading papers — TradingAgents (arXiv 2412.20138), FinMem, TradingGroup — and noticed a pattern. Every paper adds risk controls in software: a risk-controller agent that sits alongside the signal agent and says "not this trade." Smart. But software-based risk controls can be overridden, misconfigured, or simply not triggered in the moment that matters.
The original idea came from asking: what if the risk floor was somewhere no one could touch it? A BNB Chain-deployed smart contract that read on-chain price feeds and enforced max-leverage and drawdown-stop on-chain, not in a Python if-statement. We called it MacroGuard.
Scoping down, keeping honest
We eventually scoped the MVP down — no Solidity, CeFi-only on Bybit. But the honesty commitment stayed. DRIFT runs point-in-time backtests where position[t] is always shifted +1 bar before execution. The equity curve you see is the one you would have earned. The drawdown stop that halts a bot is code, not a UI toggle.
Profit Mirage (arXiv 2510.07920) is the canonical paper on backtest inflation. We read it before writing a single line of the backtester and built the shift directly into the vectorised engine.
The name stood: Deterministic Risk-bounded Intelligent Financial Trading. The point-in-time constraint and the bounded risk stop are not features — they are the product.