Strategies

ETH Single-Asset RSI Extreme Mean Reversion Long-Short on BINANCE Daily (Counter-Trend Complement to TSM, Single Instrument)

Outcome: Abandoned

EthDailyRsiExtremeMeanReversionLongShort

Outcome Summary

This strategy was pitched as the mean-reversion counterpart to an ETH TSM family — buying RSI<25 oversold dips and shorting RSI>75 overbought spikes on the thesis that ETH's daily horizon is less arbitraged than BTC 4H. The implementation was faithful and clean, but the premise didn't hold: profit factor 0.80, Sharpe -0.54, -15.2% total, and losses in five of seven years, with the short leg in particular bleeding by fading overbought extremes in an up-biased asset. The analyst abandoned it before optimization, judging it a dead, over-arbitraged premise on a major rather than a tunable parameter region, and noting that only 39 trades over six years would let optimization curve-fit to the lone 2020-2021 bull rather than find a real edge.

Strategy report

Hypothesis

Single-asset RSI-extreme mean reversion strategy on ETHUSDT.BINANCE — the COUNTER-TREND COMPLEMENT to my in-pipeline single-asset TSM family (BTC + ETH + SOL TSM). LONG when 14-day RSI drops below 25 (severely oversold, expecting bounce); SHORT when RSI rises above 75 (severely overbought, expecting pullback); FLAT when RSI is in the normal 25-75 band. This is the SIMPLEST academic counter-trend signal — Wilder's RSI (1978) at extreme thresholds has been validated as a mean-reversion trigger in equities (Lehmann 1990, Conrad-Hameed-Niden 1994) and crypto (Liu-Tsyvinski-Wu 2021 specifically test RSI signals in crypto factor literature). DELIBERATE PAIRING with the TSM hypotheses: TSM longs persistent trends; RSI MR longs/shorts at trend exhaustion points. The two mechanisms are EMPIRICALLY UNCORRELATED — TSM produces signals when trend t-stat exceeds threshold; RSI MR produces signals at trend EXTREMES (which are by definition AFTER the trend has run). Combined, they cover both behavioral regimes (persistence + reversal). ETH chosen specifically because: (1) the analyst explicitly refuted mean reversion on BTC at 4H ('over-arbitraged') but ETH at DAILY (longer horizon, ecosystem-narrative-driven swings) is structurally different and less arbitraged at this granularity, (2) ETH has more pronounced overbought/oversold extremes than BTC due to ecosystem catalyst events (DeFi summer, NFT boom, restaking thesis) that create classic 'buy the dip' setups. Single-asset scope means no multi-asset translation curse. ETHUSDT.BINANCE 1D is on the system's explicit safe-list. Helps direction quota (currently 90.0% long-only). If RSI MR validates positive Sharpe on ETH daily AND TSM validates positive Sharpe on ETH daily, the analyst has TWO complementary mechanisms on the same instrument that can be combined into a regime-adaptive portfolio.

Implements Wilder's RSI-extreme mean reversion as the counter-trend complement to the single-asset TSM family on the same instrument (ETH daily): TSM longs persistent trends, RSI MR fades trend extremes, and the two fire in different regimes (trend t-stat vs RSI extremes after the trend has run), so they are empirically near-uncorrelated and combine into a regime-adaptive pair. RSI is a bounded natural-unit oscillator, so calculate_signal returns the continuous centered value (50-RSI)/25 (varying every bar) while should_enter thresholds the raw RSI against 25/75 in its own units, never clipped, avoiding the gated-to-0.0 anti-pattern. The exit is the reversion completing -- RSI recovering through the neutral 50 line -- which captures the actual snap-back edge, with a PnL stop/take-profit and time stop as guards. Daily horizon is chosen deliberately: the analyst refuted BTC mean reversion at 4H (over-arbitraged), but ETH daily is a longer, less-arbitraged horizon driven by ecosystem-catalyst swings that produce classic oversold/overbought extremes. Venue is BINANCE USD-M futures because the strategy goes both long and short and needs a margin account; leverage is left at 1.0 (no amplification) so the leverage_set_but_unused gate does not trigger, and the 12%/8% exits keep expected per-trade edge well above the ~0.10% round-trip fee. Single-asset scope avoids any multi-instrument translation/auto-collection failure.

RSI-extreme mean reversion long-short on ETH daily is a structural net loser, not an implementation bug. The code faithfully implements the hypothesis (longs at RSI<25, shorts at RSI>75) but the mechanism has no edge: profit_factor 0.80, Sharpe -0.54, expectancy -$243/trade, and negative returns in 5 of 7 years with only the 2020-2021 early bull positive. The short leg (30 of 39 trades at 33% win rate) bleeds by fading overbought RSI in an up-biased asset, and the long leg is barely breakeven. With only 39 trades over 6 years and a negative base edge, optimization cannot find a real edge — it would only curve-fit parameters to the two lucky early-bull years. This is a dead premise (mean reversion on crypto majors is over-arbitraged, as previously flagged), not a fixable parameter region or code defect, so it is not worth 2 hours of optimization. FAILURE PATTERN: single-asset RSI-extreme mean-reversion long-short on a crypto major daily produces a net-negative, regime-concentrated result — the counter-trend short leg in particular fades overbought extremes into an up-biased asset and bleeds, and the ~39-trade sample is too thin for optimization to rescue.

Implementation

Long-short RSI-extreme mean reversion on ETHUSDT.BINANCE USD-M futures, DAILY bars, pure OHLCV single feed. Computes Wilder's 14-day RSI; goes LONG when RSI < 25 (severely oversold) and SHORT when RSI > 75 (severely overbought), flat in the 25-75 band. The signal is the continuous centered oscillator (50-RSI)/25 computed every bar; the discrete RSI thresholds gate entry. Exits when RSI mean-reverts through the neutral 50 level (longs on RSI>=50, shorts on RSI<=50), or on a 12% take-profit, 8% stop, or 15-bar time stop. Sized at 20% equity notional. leverage 1.0.

Backtest Review

Sharpe: -0.54
Total return: -15.15%
Max drawdown: 18.70%
Trades: 39
Win rate: 35.9%
Profit factor: 0.80

Implementation is correct and faithful to the hypothesis: 9 longs at RSI<25, 30 shorts at RSI>75, exits on RSI recovery through 50 plus stop/TP/time-stop

Single-asset, single-feed, no multi-instrument or stale-leg risk

Metrics are reliable and plausible (no artifact) — the result is simply negative

Net loser across the full sample: total_return -15.2%, Sharpe -0.54, profit_factor 0.80, expectancy -$243/trade

Negative in 5 of 7 years; only positive in the 2020-2021 early bull — no edge in any recent regime

Short leg (30 of 39 trades, 33% win rate) bleeds by fading overbought RSI in an up-biased asset

Only 39 trades over 6 years — too few to optimize without curve-fitting to the 2 lucky early-bull years

Mechanism matches the known over-arbitraged crypto-major mean-reversion failure pattern the analyst previously flagged

Outcome Summary

Single-asset RSI-extreme mean-reversion long-short on a crypto major daily has no edge — the counter-trend short leg bleeds by fading overbought extremes into a structurally up-biased asset, the edge is concentrated in the early-bull years, and a ~39-trade sample is too thin for optimization to rescue; counter-trend fades on majors remain over-arbitraged regardless of the longer daily horizon.

The analyst abandoned it at backtest review as a structural net loser and dead premise rather than a bug — the code faithfully implemented the RSI thresholds, but mean reversion on a crypto major is over-arbitraged (as previously flagged), the short leg systematically fades overbought extremes into an up-biased asset, and with only 39 trades on a negative base, optimization could only curve-fit to the two lucky early-bull years.

A long-short RSI-extreme mean-reversion on ETHUSDT daily perpetual futures (pure OHLCV) — the counter-trend complement to a single-asset TSM family — going long when 14-day RSI fell below 25 (oversold bounce) and short when RSI rose above 75 (overbought pullback), exiting when RSI recovered through the neutral 50 level or on a stop/take-profit/time stop.

Over only 39 trades in 6 years (9 long, 30 short) it was a net loser: profit factor 0.80, Sharpe -0.54 (CI -2.39 to 1.29), total return -15.2%, expectancy -$243/trade, win rate 36%, 18.7% max drawdown, negative in 5 of 7 years with only the 2020-2021 early bull positive; the short leg (30 trades, 33% win rate) bled by fading overbought RSI in an up-biased asset and the long leg was barely breakeven.

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