Strategies

AVAX Hyperliquid 4H: High-Recall Opportunities with a Frozen Meta-Label Filter

Outcome: Abandoned

AvaxFrozenOpportunityMetaLabel

Outcome Summary

AvaxFrozenOpportunityMetaLabel tested frozen two-stage classifiers for AVAX 4-hour trades with Binance context and a fixed holding period. After one iteration, its backtest produced 204 closed trades, a -1.994% return and Sharpe of -0.143. Backtest review abandoned it for negative net expectancy before optimization. The missing executed primary-only control means this result does not establish that the secondary classifier worsened performance or that meta-labeling generally fails.

Strategy report

Hypothesis

Derived by the Research Lead from PyRsquared's 'Meta Labeling for trading opportunities': test whether a secondary classifier improves the economic precision of a high-recall primary opportunity model. The source supplies the two-stage classification idea, not a trading strategy. Research choices are AVAX 4H bars, long/short directional candidates, regularized logistic models, chronological out-of-fold training, and a fixed holding period. AVAXUSD.HYPERLIQUID is the only traded instrument; AVAXUSDT.BINANCE supplies synchronized context features. Orders execute at the signal bar's close. Models and thresholds are trained exclusively within each historical training window and remain frozen throughout its subsequent evaluation window. Compare the filtered strategy against an otherwise identical primary-only control; neither control results nor evaluation labels may alter the registered contract.

Hyperliquid represents 10.2% of experiments versus a 20% target; long-only exposure is already 80.9%, and cross-venue research is only 7.7%. This long/short, cross-venue-informed experiment addresses those gaps without claiming to trade an arbitrage spread. Corpus concentration is empty, so AVAX's recent share cannot be verified. Four-hour bars avoid Hyperliquid's short-history sub-hourly limitation. The novel question is the incremental value of leakage-controlled meta-labeling, rather than another unfiltered indicator strategy. Options are underrepresented too, but adding rolling contracts and Greeks would introduce unrelated complexity into this source's classifier experiment.

negative_expectancy: decline optimization of this frozen AVAX 4H operationalization because its measured net per-trade return is negative and below the venue review floor across 204 trades. This verdict evaluates the Research Lead's trading specification, not PyRsquared's general two-stage classification idea. Without the executed primary-only control, it does not establish that the secondary classifier worsens performance or that meta-labeling universally fails.

Implementation

Long/short AVAX Hyperliquid opportunities filtered by a frozen secondary logistic classifier. Uses exactly synchronized Binance AVAX context, four purged chronological primary folds, training-only secondary threshold calibration, ATR-budgeted exposure capped at 25%, and a fixed 12-bar baseline holding period. Models remain frozen after the evaluation boundary.

Implements high-recall opportunity detection followed by directional correctness filtering without fitting on evaluation labels. Opportunity labels require an absolute forward move exceeding the registered cost plus margin; meta labels require the selected direction to clear the same economic hurdle. Direction follows the registered side-lookback momentum. Scaling is learned separately inside each chronological fit. The final fold calibrates the meta threshold on non-overlapping net returns; all fitted models and thresholds are then frozen. Model serialization and position counters are scalar restart-persisted attributes. The required engine-level primary-only comparison must be scheduled separately using the identical parameters and _primary_only_control=true; this submission exposes that mode but cannot schedule or report the comparison itself. Its independent shadow control records unweighted returns under the registered label-cost assumption and is diagnostic, not a substitute for engine costs or a performance report. No stop or target is added because the hypothesis specifies fixed holding. The 15% drawdown constraint remains an evaluation criterion, not a guaranteed risk bound. The requested outbox file could not be written because this session permits read-only filesystem access; the complete bound JSON is returned here.

Verification Results

Run a separate primary-only engine evaluation with identical historical boundaries, parameters, sizing, and execution costs before assessing whether the secondary classifier improves economic precision.

The shadow control records unweighted returns using a fixed label-cost deduction. It does not reproduce engine sizing, commissions, impact, or equity accounting. The primary-only execution mode exists, but the supplied verification contains no paired engine comparison, and no consumer of primary_only_control_required was found in src/ or tests/.

The sandbox reports -0.210% average net trade return and impact costs equal to 162% of gross profit. Full-history evaluation should compare the filtered strategy with the engine-executed primary-only control using identical costs and sizing; these smoke-test results do not establish a correctness failure.

Backtest Review

Sharpe: -0.14
Total return: -1.99%
Max drawdown: 10.83%
Trades: 204
Win rate: 50.5%
Profit factor: 0.96

204 closed trades over 556 days provide a functioning evaluation sample. Inspected trades show AVAX-only long/short positions, approximately two-day holds and varying sizes, consistent with the implementation.

Accounting reconciles, all positions are closed, and synchronized bar audits report no missing bars.

Net average trade return is -0.225% versus the Hyperliquid review floor of +0.15%; PF is 0.958, Sharpe -0.143 and total return -1.994%. This baseline does not warrant optimization.

Modeled impact is $8,730, reported as 184% of gross profit, with estimated capacity of $29,509. QA's economic concern is borne out.

The required engine-executed primary-only control is absent from the supplied report and detail file. Incremental benefit from meta-labeling remains unmeasured; the shadow ledger omits actual execution costs, funding and capital-relative sizing.

Forecast IC is -0.096 with t=-1.3: supporting evidence provides no demonstrated predictive benefit, but does not prove the general mechanism absent.

Outcome Summary

A future iteration should demonstrate positive net expectancy and include an engine-executed primary-only control before attributing results to meta-labeling.

Backtest review abandoned it because negative net expectancy fell below the +0.15% venue review floor; optimization and later review stages were not reached. The required executed primary-only control was absent, leaving the filter's incremental benefit unmeasured.

It tested whether a frozen secondary logistic classifier could improve high-recall AVAX 4-hour long/short opportunities on Hyperliquid, using Binance AVAX bars as context.

The backtest recorded 204 closed trades over 556 days, returning -1.994% with Sharpe -0.143 and maximum drawdown 10.831%. Average net trade return was -0.225% and profit factor was 0.958.

Analysis

Meta Labeling for trading opportunities

Backtest and paper results are hypothetical. Trading involves risk of loss.