Flat probabilities: what zero signal actually looks like

RSI. MACD. Stochastics. ADX. ATR breakouts. Momentum. The overwhelming majority of retail trading education is built on the premise that these indicators, alone or in combination, carry information about the next move. It is the most widely believed — and least honestly tested — hypothesis in the industry. Testing it properly was the first thing we did after auditing away our ensemble illusions (E-01).

The test rig matters more than the test, so here is the full spec of walk_forward.py, the honest harness this project runs on:

• Clean data: 69,000 5-minute XAUUSD bars fetched natively from the broker API — cross-validated against an older export ($0.05 median candle difference once a hidden 3-hour timezone offset was found and corrected; yes, that pathology too).
• Stationary features only — indicator values and returns, never raw price levels a model could use to memorize eras.
• Per-feature scalers fit on training data only — the exact anti-leak discipline whose absence produced E-01.
• Triple-barrier labels — up-barrier, down-barrier, time-out — so “direction” means something tradeable, not a sign flip.
• Chronological walk-forward folds, decision thresholds chosen on a validation window and applied untouched to the out-of-sample test window.
• Real economics on every simulated trade: $100 per $1 move per lot, $0.35 spread, $7 round-turn commission.

XGBoost, trained on the classic indicator set across rolling folds, produced out-of-sample predicted probabilities whose 5th-to-95th percentile range was [0.498, 0.503].

Sit with that number. Across thousands of unseen bars, the model's belief about direction never strayed more than three-tenths of a percent from a coin flip. Not “weak signal” — no signal. Out-of-sample accuracy landed at or below the majority-class baseline. Forcing trades anyway lost between $604 and $6,803 per fold. There was nothing there to trade.

Our read, stated as interpretation rather than theorem: every classic indicator is a deterministic transform of past prices. If the raw price series at this frequency carries no exploitable structure beyond the cost wall — and both this experiment and our order-book work (E-02) point the same way — then no recombination of that series manufactures information that isn't in it. Indicators repackage; they cannot create. Whatever genuine edge these tools carried decades ago has had every basis point arbitraged out by the thousands of funds running the same arithmetic faster.

The two findings corroborate each other from opposite ends of the data spectrum: quote-level microstructure topped out at a ceiling too thin to pay for itself, and bar-level indicators didn't reach a ceiling at all. Price-derived features, at retail costs, contain no standalone directional edge.

Here is where most projects go wrong, and where ours nearly did: faced with a flat result, the reflex is to tune — different barriers, different lookbacks, more features, until some configuration finally shows green. With enough configurations, one always does. That green is not discovery; it is selection. Our working rule, written down after this experiment and enforced since: never tune until profitable — that is overfitting a signal that has been shown not to exist.

A pipeline that can output “flatly, honestly, nothing” is the single most valuable piece of equipment in this lab. Most pipelines in this industry cannot output that answer, which is exactly why the industry publishes so many edges and so few audits.

No out-of-sample signal at all — the cleanest kill in the log. Killed not by costs (it never earned the right to face them) but by the out-of-sample split itself. What survived is the harness: the same walk-forward rig went on to test — and kill — every subsequent strategy in this journal, and to validate the one model that passed. Every number on this site descends from the discipline this experiment forced on us. Read the full standard in our methodology.