Elliott wave Fibonacci retracement strategy: define a test before scoring it

Define the sample and anchors

Choose the instruments, timeframes, observation period and eligibility rules in advance. State which wave hypothesis and anchors the ratio depends on. Keep cases that later become invalid or unresolved, so the aphigher-degree wave result is not improved by removing inconvenient examples.

Make the decision rule reproducible

Specify whether the event is a touch, a close or a later structural condition. Define when information becomes available and how a case is abandoned. Two reviewers should be able to apply the written rule to the same preserved data without guessing your intention.

Separate a signal from execution

If the study includes simulated trades, define the assumed execution, spread, fees and treatment of gaps. Do not assume a historical level could always have been traded at exactly that price. These assumptions belong to the protocol and can materially change the interpretation.

Evaluate fresh observations

Keep the cases used to develop the idea separate from later evaluation cases. Report sample size, unresolved cases and limitations alongside any eventual measures. No such performance study is provided here, and correct Fibonacci arithmetic alone does not demonstrate a trading advantage.

A teaching timeline for evaluation

After development on indices 1–100, freeze the initial protocol at 100; separate evaluation uses 101–150. All figures and cases are invented. In A, completed bar 110 closes above 103, an assumed purchase 104 occurs at 111 and sale 112 at 116. With q=24, round-trip commission 0.4 per unit and fixed costs 2, net result is 24×(8−0.4)−2=180.4. In B, the condition occurs at 135, purchase 104 at 136 and gap exit 99 at 138, giving −131.6.

In C, the condition occurs at 140 and purchase 104 at 141. Bar 142 has O 104, H 113, L 100, C 108: both stop 101 and target 112 are touched, but OHLC does not reveal their order. Retain C as unresolved, without deleting it or choosing the favorable outcome. A+B=48.8 sums only the two known cases, not the complete cohort. With the specified fills, do not deduct the 0.6 allowance again. This illustrates information accounting, not a measured market or product backtest.

• Bailey et al.: backtest overfitting
• Investor.gov: an observed level is not an execution price

Walk-forward with label availability

In a separate example, every label matures exactly two observations after anchor i. Round 1 trains on anchors 1–20, whose last labels are available after 22; freeze after 22, exclude anchors 21–22, and test 23–26 is fully scored only after 28. Under a preselected expanding schedule, round 2 trains 1–26 available by 28, gaps 27–28, freezes 28 and tests 29–32 with final outcomes available by 34.

Earlier test 23–26 can enter later training only once its outcomes are known; it is not newly untouched OOS. Select settings and fit preprocessing on the corresponding training set, never future tests. Gap 2 is sufficient only for this label horizon and does not replace overlap/position analysis. Do not report the final test outcome at 32 before labels mature at 34; dependent real observations need separate uncertainty assessment.

• scikit-learn · TimeSeriesSplit
• NIST · Sample estimates and uncertainty

Sample expectancy after commission

Use a different complete teaching cohort of 10 closed trades, not the walk-forward results above. Eight have gross+1.5 and two gross−5 under assumed fills; round-trip commission is 0.5 per trade. Net winners gain 1, net losers lose 5.5. Gross totals+2, fees 5, net−3; sample mean Ê=−3/10=−0.3 account-currency units per trade.

With no zero outcomes, Ê=p×W−(1−p)×L=0.8×1−0.2×5.5=−0.3, where W/L are average net gain and absolute net loss. Do not subtract commission twice or add slippage again if already reflected in the adopted fills. This is the mean of supplied numbers, not an established future expectation or Elliott validation. A real protocol preserves the cohort, sizes, unresolved cases and all material costs.

• NIST · Sample estimates and uncertainty

An 80% win rate with a negative result

In the same separate cohort, a net winner means strictly net>0 after the specified commission. Eight of 10 qualify: win rate 80%; two have net<0, with no zero outcomes. Net still totals−3 and averages−0.3 because each 5.5 loss exceeds each 1 gain. Correct wave labels or touched levels define different metrics.

Predefine the denominator and treatment of zero, unfilled and unresolved cases; do not delete poor results. Observed 80% is not a future win probability or profitability confirmation. Win rate cannot replace outcome size, drawdown, costs and time dependence; real uncertainty requires separate assessment.

• NIST · Sample estimates and uncertainty

Sources and limits

• Fibonacci: retracements, projections and checks.
• What a wave count can and cannot establish.