Paper Trading vs Live Trading for AI Strategies

Paper trading lets an AI trader build a decision record without exposing real funds. It removes execution pressure from the user and gives the system room to collect enough trades for a fair review.

Live trading adds fees, slippage, liquidity limits, emotional pressure, and operational risk. That is why a strong paper record should be treated as a candidate for deeper review, not as proof that live results will match.

Paper engines usually match decisions against recorded quotes or an exchange feed, update a virtual balance, and apply configured fee and fill rules. That isolates strategy logic and makes repeated evaluation inexpensive, but the closer those rules mirror order types, latency, funding, and partial fills, the more informative the record becomes.

Simulation cannot reproduce queue position, market impact, rejected orders, API interruptions, or the temptation to override a losing system. Historical liquidity may also disappear precisely during stress. Conversely, overly pessimistic fill assumptions can reject a viable idea, so paper results depend heavily on transparent modeling choices.

Consider a strategy that buys at a displayed price of $100 and exits at $101. A frictionless simulator records 1%, but a 0.10% fee on each side plus 0.20% combined slippage reduces the gain to roughly 0.6% before funding; repeating this difference across frequent trades can reverse an apparently profitable result.

Move beyond paper only after enough trades span more than one regime and results remain acceptable under harsher cost assumptions. Begin with capital small enough that execution errors are tolerable, define a stop condition for divergence from the paper record, and remain in simulation if the system cannot be monitored or safely disabled.

Aigentra's default paper configuration starts accounts at 10,000 USDT, records maker fees at 0.02%, taker fees at 0.05%, and applies a 0.01% adverse slippage assumption to configured market-like exits. These are engine defaults, not promises about any exchange. Each event should display the fee and maker or taker role actually stored for that simulated fill. Equity-based return includes fees already deducted, while funding is included only when the engine explicitly records it.

The missing live variables are just as important as the modeled ones. A paper engine cannot know an account's real queue position, market impact, exchange throttling, liquidation engine behavior, or whether a user would interrupt the strategy during stress. A useful promotion test therefore compares small live fills with paper events by timestamp, requested price, executed price, fee, rejected quantity, and delay. Stop the test when divergence exceeds a written limit rather than rationalizing each miss after it occurs.