ATOM 4H Parabolic SAR Trend-Following Long

Hypothesis

Fundamentally unprofitable and not optimizable: total_return -42.7%, profit_factor 0.898 (<1), negative expectancy, avg_trade_return_pct 0.137% below the 0.15% futures fee floor, and max_drawdown 68.8% (CI up to ~100%) far exceeds the 50% hard-abandon limit. Losses persist across every recent year including 2025 (-50%), so the trend edge is absent in-regime, not merely decayed. The positive Sharpe is a skew/kurtosis artifact of a handful of outlier days (skew 4.7, kurtosis 63) on a losing equity curve. This is the single-name OHLCV trend-follower pattern with zero survivors; retuning AF constants cannot conjure an edge a naked SAR flip on one alt-perp does not have. Failure pattern: fee_edge / no-edge single-name trend follower with catastrophic drawdown.

Implementation

Long-only Parabolic SAR (Welles Wilder 1978) trend follower on ATOMUSDT Binance USD-M perpetual futures, 4-hour bars, pure OHLCV with zero supplementary-data dependency. Each bar recomputes the SAR over the buffered history using canonical Wilder acceleration factors (start 0.02, step 0.02, max 0.20). Enters LONG when the SAR sits below price (uptrend) and exits to flat when the SAR flips above price (downtrend) — the parabolic SAR is itself the trailing stop, so no separate stop/target is layered on. Positions are risk-sized at 2% of equity over the price-to-SAR stop distance, capped at 1x equity notional (leverage 1.0).

Verification Results

CLEAN RESTART 2026-09-04 — this run's verdict history and learning records were removed and it was restarted from verification. Its previous abandonment came from the pipeline, not from the market: the Layer-2 harness mis-bound @staticmethod helpers (fixed), QA issued terminal performance verdicts on an unoptimized smoke test (removed — QA now judges correctness only), and sandbox timeouts came from backtest-slot starvation (fixed). The hypothesis and the strategy code are unchanged. Verify the code on its merits; performance is decided later by the full backtest and the optimizer.

Persist up/af/ep/sar across bars and update only on the new bar (O(1)), removing the sliding-seed subtlety.

The SAR is recomputed from scratch over the ENTIRE buffer every bar (O(n²)), re-seeding up=True/sar=lows[0]/ep=highs[0] at the buffer start each time. Because the buffer slides, the streaming SAR is not strictly path-identical to a once-initialized continuous SAR. Immaterial in practice — parabolic acceleration erases the seed within ~10-20 bars, well inside the buffer — but an incrementally maintained SAR would be faster and path-exact.

Lower the notional cap or floor the SAR stop distance as a % of price so early-trend entries aren't systematically maxed; let the optimizer choose.

Risk sizing (2%*equity/(price−SAR)) drives most entries to the 1x cap: right after an up-flip the SAR is close to price, so the stop distance is tiny and size clamps to equity/price. The book enters at full 1x notional at trend starts, producing the ~50% avg position and deep drawdowns. Not a correctness bug (no leverage blowup), but the small-risk-per-trade intent is largely overridden by the cap.

Backtest Review

Sharpe

0.53

Total return

-42.72%

Max drawdown

68.85%

Trades

606

Win rate

38.1%

Profit factor

0.90

Clean, verification-safe canonical implementation (Parabolic SAR, pure OHLCV, no supp-data dependency); 606 trades is an adequate sample

New symbol/mechanism as intended — genuinely distinct family, so the negative result is informative

Loses money outright: total_return -42.7%, profit_factor 0.898 (<1), negative expectancy -$104/trade, win rate 38%

avg_trade_return_pct 0.137% is below the 0.15% Binance USD-M fee floor — gross edge does not clear costs

max_drawdown 68.8% (CI 79.6-99.96%) — well past the 50% hard-abandon threshold

Negative in every recent year (2022 -21%, 2023 -18%, 2024 -11%, 2025 -50%); no edge in the modern regime, contradicting the regime-decay-mitigation thesis

Positive Sharpe 0.53 is an artifact of extreme positive skew (4.7) / kurtosis (63) — a few outlier days on a losing curve; Sharpe CI [-0.07, 1.09] straddles zero

Iteration History

data_unavailable: backtest data for ATOMUSDT.BINANCE could not be collected after 5 retries: - bars (bar_type=ATOMUSDT.BINANCE-4-HOUR-LAST-EXTERNAL): got 0, needed 100, or 60+ days span The data layer (not the strategy) is the blocker — re-evaluate this hypothesis once collection is fixed / the symbol is backfilled.

Outcome Summary

This hypothesis was carefully designed around prior failure patterns — pure OHLCV with zero supplementary-data dependency, a long-only trend-following mechanism, the proven 4H + Binance USD-M combination, and a deliberately new symbol (ATOM) and indicator (Parabolic SAR) for diversification. The implementation reached the backtesting stage, but the engine retrieved 0 bars for ATOMUSDT.BINANCE 4-hour data against a required minimum, and after 5 retries the run was abandoned as a data-availability problem. Because the blocker was the data layer rather than the strategy, no backtest, optimization, or analyst evaluation ever occurred, and the abandon note explicitly flags it for re-evaluation once collection is fixed or the symbol is backfilled. The strategy's edge therefore remains untested rather than disproven.

Verify that the target symbol's bar data is actually available/backfilled before committing a hypothesis to the pipeline; a deliberately novel symbol choice is wasted if the data layer cannot supply its history.

It was abandoned at the data/backtesting stage due to a data-layer blocker, not a strategy fault — backtest data for ATOMUSDT.BINANCE could not be collected after 5 retries — so the backtest-review, optimization, analyst, and risk stages were never reached.

A long-only, single-instrument, OHLCV-only trend-following strategy on ATOMUSDT Binance USD-M perpetual 4-hour bars using a canonical Welles Wilder Parabolic SAR (AF start 0.02, step 0.02, max 0.20) — entering long when the SAR flips below price and exiting flat when it flips above, with the SAR itself as the trailing stop.

No performance exists: the backtest never ran because the required ATOMUSDT.BINANCE 4-hour bar data could not be collected (0 bars retrieved against a needed minimum of 100 / 60+ days span). No metrics, optimization, or analyst review were produced.