Extrema Signals Looked Smart, Then Lost Money on Binance Spot

The Practical Question: Can Extrema Prediction Pay After Binance Spot Costs?

Ayoub Jadouli's paper asks a clean desk question: can candle-based machine learning models predict local crypto lows or highs well enough to make money on Binance Spot after assumed paper-trading costs?

That distinction matters. A model can rank future local minima well and still fail once it has to trade at an executable next price, hold inventory, pay costs, avoid churn, and survive bad exits. The paper is not trying to prove that machine learning cannot work in crypto. It is narrower than that, and more useful. It audits a set of specific Binance Spot timing policies and asks what still holds after chronology, costs, execution timing, search history, and artifact checks.

The answer is mostly no. In the author's own operating language, every decision remains NO_TRADE.

What Was Tested: Daily Selectors, Local Min/Max Policies, and an OHLCV-Only Gurgul Adaptation

The tested universe is candle-based crypto timing, mainly using Binance Spot data. The paper covers several model families, but the strongest claims are tied to artifact-backed runs rather than older narrative results.

The first key family is an unchanged ten-pair mandatory-daily selector. It has to pick a trade each daily cycle, which makes it a useful stress test. The policy cannot simply stand aside when signal quality looks poor.

The second family is a set of local-extrema policies. One tries to buy around predicted local minima. Another treats predicted local maxima as sell-to-cash or re-entry timing signals. These are closer to the economic story a trader would want: buy near local lows, avoid holding through likely local highs.

The third family adapts the idea behind Gurgul, Lessmann, and Härdle's paired-extrema work, but only with OHLCV inputs. That caveat matters. The original study used richer information sets, including financial, blockchain, GitHub, search, macroeconomic, and text or social variables. Jadouli's version keeps the paired local-minimum and local-maximum framing, but not the full data design. It is an adaptation, not a reproduction.

The paper also discusses older One4All and slower-rotation experiments, mostly as audit context. It separates quantitative evidence from descriptive history, which is the right move. A strategy with missing raw directories should not sit in the same bucket as a reproducible later-period ledger.

The Core Result: Strong Ranking Metrics, Negative Trading Ledgers

The headline result is a familiar one for ML trading research: ranking metrics looked strong, trading ledgers did not.

The unchanged ten-pair mandatory-daily selector lost 6.72 percent over 19 July cycles at an assumed 31 bps completed-cycle cost. It had 3 wins and 16 losses. That is not a close call, even allowing for the small sample.

The local-extrema policies were also negative. The validation-selected local-minimum policy returned -1.79 percent in the short July evaluation. The local-maximum sell-to-cash and re-entry policy underperformed continuous holding by 2.80 percent. Their gross mean advantages were 11.11 bps and 12.21 bps, below even the 21 bps stress cost. In plain terms, the measured edge was smaller than the cost hurdle for trading it.

The OHLCV-only Gurgul-inspired adaptation is the most instructive case. It reached ROC AUC of 0.874 for minima and 0.896 for maxima. Those numbers would look good in a prediction paper. But average precision was only 0.134 and 0.116, a warning sign for rare events. The resulting policy lost 44.30 percent over seven cycles, versus -41.20 percent for buy-and-hold.

So the model could rank rare extrema better than chance, but not in a way that created usable spot timing value under the simulator. This is exactly why precision, event prevalence, and policy P&L need to sit beside AUC.

A related BTC-USDT walk-forward study cited in the paper found a similar split: naive sign policies failed once 10 bps costs were imposed, while cost-aware magnitude filters restored profitability in selected configurations. That comparison is useful because it points to the real issue. Architecture alone is not the trade. The trade is the whole decision rule, including abstention and turnover.

Why the Evidence Got Weaker Under Audit: Leakage, Same-Close Entry, Search, and Missing Artifacts

The paper's best feature is not the model design. It is the audit.

An earlier One4All "30-day holdout" was downgraded after review. The dates had already influenced prior architecture work, so the period was not clean. The four-hour outcome horizon was not purged at split boundaries, creating leakage risk. The simulator used same-close entry, which lets a strategy trade at a price that would not be known after the signal is formed. The raw result directories were also absent.

Each issue is common in crypto ML backtests. None is exotic. Together they are enough to move a result from evidence to anecdote.

The paper also admits extensive predecessor search. That matters because many model families, thresholds, labels, and policy variants were tried before the later negative runs. A frozen later-period test is more credible than an in-sample chart, but it is still conditioned on the research path that created it. The author does not claim a formal data-snooping correction. Instead, he reports the search history and refuses to promote the strategy.

That restraint is valuable. A weaker claim, honestly stated, is more useful than a polished equity curve with missing artifacts.

What a Desk Should Take From It: Treat Extrema Scores as Raw Material, Not a Strategy

The practitioner takeaway is not that extrema prediction is worthless. It is that extrema scores are raw inputs. They need a separate economic layer.

A desk would want next-bar execution, purged walk-forward splits, explicit costs, turnover limits, calibration-only threshold selection, and the right to abstain. It would also want to know whether the signal survives outside OHLCV features, across venues, and under live-like order constraints.

The paper gives no reason to deploy these tested policies. It does give a useful warning: high AUC on rare local lows and highs can still map to a losing ledger. The next useful experiment is not another prettier classifier. It is a smaller policy test where every accepted trade clears the cost hurdle before the order is even considered.

For each walk-forward month:
  1. Train on prior data after purging/embargoing label horizons.
     - Classifier A: probability current bar is near a local minimum.
     - Classifier B: probability current bar is near a local maximum / adverse exit risk.
     - Regressor: next executable open to 4h-later open return in bps.
  2. Calibrate classifier probabilities on the calibration fold only.
  3. Select entry thresholds on calibration data subject to:
     - positive calibration net return after costs
     - positive mean trade expectancy
     - max drawdown below required gate
     - minimum calibration trade count
  4. If the policy gate fails, do not trade that test fold.

On each completed 5-minute bar during the test fold:
  Build OHLCV-only features using past data only.
  Predict:
    p_minimum, p_maximum_risk, expected_4h_return_bps
  Compute predicted_net_edge = expected_4h_return_bps - 31 bps primary round-trip cost.

Entry rule when flat / capacity available:
  Enter long at next resampled bar open if:
    p_minimum >= selected minimum-probability threshold
    predicted_net_edge >= selected net-edge threshold
    p_maximum_risk < 0.35
    next execution bar exists
  If multiple symbols qualify, choose highest edge minus cost.
  Size as equal-weight capped position, max 10% NAV per symbol, max 3 concurrent positions.

Exit rule for each open position:
  Exit when any condition is true:
    stop loss hit: -75 bps
    profit target hit: +120 bps
    timeout reached: 48 primary 5-minute bars
    p_maximum_risk >= 0.55
    expected net edge <= 0 bps
    symbol/execution price unavailable
  Evaluate stop/target using resampled bar high/low; if same-bar tie, assume adverse stop first.
  Apply half round-trip cost per side, including terminal liquidation.