How We Measure Accuracy

Last updated: June 11, 2026

MidasFlow measures signal accuracy by logging every emitted signal and evaluating its outcome at fixed time horizons after emission. A pump signal counts as successful when price reaches the alert's projected target zone before hitting its invalidation level. The 72–85% consensus accuracy quoted on our homepage is the share of signals where the multi-agent consensus direction matched the realized move, computed over rolling windows — it varies with market regime and is never a promise of returns.

1. What counts as a signal

A signal is an alert actually delivered to users, with its score, levels and timestamp logged at the moment of emission. Nothing is added or edited retroactively: if the system was wrong, the record stays wrong. Internal experiments and shadow tests are tracked separately and never mixed into public accuracy figures.

2. How outcomes are evaluated

Each signal is re-checked at fixed horizons after emission (minutes to hours, depending on signal type). The outcome is classified as:

• Success — price reached the projected target zone before the invalidation level;
• Failure — the invalidation level was hit first;
• Fade — neither happened within the evaluation window.

Fades are reported separately rather than silently counted as wins — a practice we consider non-negotiable for honest statistics.

3. What "72–85% consensus accuracy" means

Multiple agents evaluate the same market simultaneously. When their weighted consensus points one way, we track whether the market followed. Over rolling evaluation windows this consensus has historically landed in the 72–85% range depending on market conditions — trending markets sit near the top of the range, choppy regimes near the bottom. It is a directional-agreement metric, not a P&L metric: a correct direction does not guarantee a profitable trade after fees, slippage and timing.

4. Agent score cards on the homepage

The per-agent percentages (e.g. "SMC Trader 96%") are backtest-based estimates on historical data within each agent's specialty domain, recalibrated periodically. They describe how well each agent's logic fit past market behavior — they are not audited live trading results and not a forecast of future performance.

A public monthly Signal Accuracy Report with verified production numbers is in preparation and will be published on this site.

5. Known limitations

• Regime dependence. Accuracy degrades when market structure shifts abruptly; models are retrained, but there is always a lag.
• Early detection fades. Roughly 30–40% of early cross-exchange warnings do not develop into a full move — documented in our Early Detection guide.
• Liquidity. On small-cap symbols, slippage can consume the edge a correct signal provides.
• Past ≠ future. Every figure on this site describes historical behavior.