May 2026 — Strategy Metrics¶
Data snapshot: 2026-05-30 23:07 UTC (live/stag). Intra-month live snapshot 05-27, paper cohort 05-26.
Active Live Strategies (05-30)¶
| Strategy | Status | P&L | Trades | Volume | Win Rate | Margin (bps) | Avg Hold |
|---|---|---|---|---|---|---|---|
| MMTV_v15_eth_invgate | ACTIVE | +$0.15 |
494 | $23,496 |
59% | +0.06 | 1.7m |
| MMTV_v15_eth_tape | ACTIVE | +$0.14 |
1809 | $84,450 |
60% | +0.02 | 1.4m |
| MMTV_v15_eth | ACTIVE | -$0.47 |
2377 | $107,110 |
61% | -0.05 | 0.9m |
| MMTV_v15_btc_invgate | DISABLED | -$1.67 |
351 | $17,388 |
54% | -0.96 | 1.9m |
| MMTV_v15_btc_tape | DISABLED | -$1.07 |
982 | $48,673 |
57% | -0.22 | 1.7m |
| MMTV_v15_btc | DISABLED | -$1.80 |
1343 | $66,679 |
61% | -0.27 | 0.9m |
| MR_v19_ETHUSDT | ACTIVE | +$0.14 |
1 | $50 |
100% | +27.2 | 23.9m |
| MR_v19_BTCUSDT | ACTIVE | +$0.21 |
1 | $50 |
100% | +41.1 | 32.9m |
| MR_v19_tight_sr_eth | ACTIVE | +$0.13 |
1 | $50 |
100% | +25.6 | 23.9m |
| MR_v19_tight_sr_btc | ACTIVE | +$0.17 |
1 | $50 |
100% | +34.1 | 32.3m |
| TF_v18_ETHUSDT | ACTIVE | -$0.79 |
1 | $50 |
0% | -157.4 | 12.9h |
| TF_v18_BTCUSDT | ACTIVE | -$0.51 |
1 | $50 |
0% | -102.8 | 17.6h |
| TF_v18_breakout_eth | ACTIVE | -$0.85 |
1 | $50 |
0% | -170.3 | 13.0h |
| TF_v18_er_capped (eth/btc) | ACTIVE | $0.00 |
0 | — | — | — | — |
| StatsArb_v19_ETH_BTC | ACTIVE | +$0.05 |
13 | $750 |
46% | +0.69 | 1.2h |
Read: MM v15 ETH variants marginally positive at scale (invgate/tape), but the structural-viability memo shows
this is within noise of the adverse-selection ceiling, not edge. All BTC MM variants negative + disabled. MR v19 is
the live candidate — all 4 variants +26 to +41 bps, 100% win; small n but uniformly positive and not falsifiable
in our backtest (held, not killed). TF v18 negative on every variant that traded. StatsArb near-flat.
The Structural Verdicts (the month's real output)¶
MarketMaking — unviable on BTC/ETH spot¶
| Hypothesis (v9–v15) | Verdict |
|---|---|
| Spread/edge tuning | Dead — wider quotes didn't cut fills |
| Latency / speed | Falsified — adverse drift flat −1.7 bps independent of rest-time |
| Micro-price signal | Dead — bounded ±0.25 bps on 1-tick book |
| TV-divergence taking | Dead — backtest at 5 bps taker → 0 of 1812 clear; at the real 1.75 bps taker only ~1-3.6% clear, same adverse mechanics |
| Cross-venue Binance→OKX lead | Real but negligible — IC×σ ≈ 0.04 bps, ~40× under the 1.75 bps taker hurdle |
| Vol / volume-spike / tape gates | All failed |
Cause: per-fill P&L = maker rebate − adverse selection ≈ +0.75 − 1.7 = −0.95 bps. We do earn the OKX maker
rebate (−0.75 bps), but adverse selection beats it ~2:1 on a 0.01-0.05 bps spread with no queue priority. The v9-v15
cohort modeled maker_fee_bps=0 (rebate ignored) → quoted ~2 bps off TV; a rebate-aware config (~1 bps quote) is an
untested operating point that probably doesn't flip the verdict. Structural, not a tuning gap.
Spot Direction — verdict is NOT uniform (TF dead; MR held)¶
| Hypothesis | Verdict |
|---|---|
| TF SMA crossover (v14-18) | Dead — net-negative 4 regime windows, even at zero fees |
| TF Donchian breakout | Dead — ETH down-trend −16.51 / 43 trades / 35% win |
| TF exit tuning (SL/trail) | Dead — baseline is local optimum, loss-tail structural |
| TF 1h timeframe | Dead — loss-tail timeframe-invariant |
| MR 5m (v17-19) | NOT dead — held for live evidence. Kline backtest blind to 2s intra-bar entries (uninformative offline). Live: all 4 v19 variants +26 to +41 bps, 100% win, ~25-30min holds. Small n (~3 independent events × 2 symbols) |
| Kline flow features | Weak — IC < 0.04, inconsistent across symbols |
| Short-horizon reversal | Regime-luck — IC 0.05 isn't a business |
| Multi-day reversal | Regime-luck — inconsistent quarter signs, maxDD to −163% |
| Intermediate (2-4h) reversal | Dead full-history — 360d Sharpe +1.31 → −0.42 over 2.5yr |
| AdaptiveDirectional | Dead by transitivity on the TF arm — routes a dead edge + an unprovable one |
Cause (TF/breakout/reversal): inverted R:R + regime-dependence. Every in-sample edge shrank to zero/negative under full-history + non-overlapping + net-of-cost + per-quarter validation.
MR is the exception: its 2-second intra-bar entries are invisible to a 5m-kline backtest, so "offline shows no edge" is structurally uninformative — not a verdict. Live is the only ground truth and it's uniformly positive so far, on a sample too small to confirm. Held, not killed (PR #93).
May Cohort Aggregate P&L¶
Sum of all 84 strategies created in May, back-filled from the two in-May snapshots (frozen final value for disabled strategies @05-27; 05-30 23:07 value for the 17 still-active). The drifting June live snapshot is bypassed. No incident windfall in May, so this needs no asterisk (unlike Mar/Apr). Apples-to-apples with April's −$8.87 "April-created total".
| Split | P&L | Count |
|---|---|---|
| Active at month-end | -$2.17 |
17 |
| Disabled / ERROR / KILLED (frozen) | -$50.11 |
67 |
| May-created cohort total | -$52.28 |
84 |
| Family | P&L |
|---|---|
| MarketMakingTrueValue | -$43.37 |
| TrendFollowing | -$4.23 |
| StatsArb | -$2.98 |
| StatsArbOrderflow | -$1.94 |
| MeanReversion | +$0.24 |
Read: $43 of the $52 cost is MarketMaking — the v6→v15 churn, dozens of variants trading large volume (one
MMTV_v6 instance alone did $105K volume for -$15.69). That -$43 is the empirical price of the MM-unviability
verdict: thousands of tiny adverse fills, each a data point proving spread < adverse cost. May's iteration cost was
~6× April's, concentrated in the MM investigation. The non-MM families netted ~-$9, in line with April.
Method caveat: 4 MMTV_v14 instances disabled between 05-27 and 05-30 use the 05-27 value (true final slightly more
negative); net effect < $1. Active portion carries ~1 day of residual drift.
Cumulative (all-time, through May)¶
| Category | Note |
|---|---|
| Active live | Near-flat; MM-ETH marginally +, MR tiny sample, TF negative |
| May cohort | -$52.28 — $43 of it the MM-viability investigation (see above) |
| All-time | Net-negative excluding the Mar/Apr incident windfalls |
| Phase 1 status | Not met. Current-instrument-set families verdicted dead (MR held). |
Strategy Type Evolution (Mar → Apr → May)¶
| Type | Mar | Apr | May | Verdict |
|---|---|---|---|---|
| MarketMaking | -$3.78 ERROR |
-$0.71, $63K vol |
ETH marginal +, BTC − | Structurally unviable (spot BTC/ETH) |
| MeanReversion | +$2.29 |
+$1.59 |
v19 all-green, small n | HELD — live-positive, unfalsifiable offline |
| TrendFollowing | -$0.44 |
+$0.90 |
net-negative all variants | Dead — structural R:R |
| StatsArb | +$1.02 |
-$3.43 |
near-flat | Parked |
| AdaptiveDirectional | — | paper, control-beat (small n) | dead by TF arm | Routes a dead + an unprovable edge |
The Apr "MM volume validation" story dissolved under May's rigorous backtest — it was within the adverse-selection ceiling, not edge. MR is the one Apr thread that survived: its offline picture is uninformative (backtest blind to its entries), and live stayed positive. The lesson cuts both ways — the methodology kills in-sample artifacts (MM) and protects signals the backtest structurally can't see (MR).