Stream 9 Wave 2 · Decision Brief

ATLAS Sizing Cutover — Conservative · Balanced · Aggressive

Generated 2026-05-08 · Cohort: 15 canonical strategies (3 gated bundles + 12 diagnostic-grade pins) · Pipeline commit 2b47a4f
In this brief
  1. Executive summary
  2. The three modes — at a glance
  3. Why deployed sizing must change
  4. Per-strategy Kelly inputs
  5. Conservative — pros, cons, when to pick it
  6. Balanced — pros, cons, when to pick it
  7. Aggressive — pros, cons, when to pick it
  8. Strategy-level need-to-knows
  9. Reading the Monte Carlo result correctly
  10. Leverage & cross-margin context
  11. What this rerun does NOT model
  12. Recommended path forward

1 · Executive summary

The deployed BTC and ETH portfolio sizing was tuned on a sizing investigation that consumed gate-broken trade lists — i.e. trades the strategies wouldn't actually have taken under correctly-running EMA-22 gate logic. Wave 2 reruns the full pipeline on the canonical pins (the trade lists each strategy actually produces under gate-active conditions) and produces three candidate configs.

All three modes return 0.0% liquidation rate under 200-trial Monte Carlo perturbation at ±30% on the per-trade returns. Liquidation safety is not the gating decision. The gating decision is how aggressively you want to capitalise on the canonical edge given that you no longer trust the deployed sizing to reflect reality.

Bottom line
Conservative is "stop the bleeding" — protect against the chance that some strategies have genuinely lost edge. Balanced is "trust Kelly, don't over-extrapolate" — the half-Kelly maths is a known-good frontier and Balanced sits roughly there. Aggressive is "let the canonical winners run wide" — high-trade-count strategies (f1_follow_short, s1_gated) get sized above current deployed levels.

2 · The three modes — at a glance

Mode A
Conservative
Total notional107%
Active strategies11 / 15
Base terminal+42.7%
Max DD (base)-0.6%
MC median+42.1%
MC P10 → P90+34.9% → +50.6%
MC liquidation0.0%
Mode B
Balanced
Total notional433%
Active strategies12 / 15
Base terminal+251.9%
Max DD (base)-2.3%
MC median+245.5%
MC P10 → P90+191.0% → +321.1%
MC liquidation0.0%
Mode C
Aggressive
Total notional1083%
Active strategies13 / 15
Base terminal+1481.6%
Max DD (base)-5.3%
MC median+1406.3%
MC P10 → P90+961.6% → +2187.0%
MC liquidation0.0%

"Notional" = sum of per-strategy notional %. Cross-margin compresses required margin so 433% notional ≠ 433% margin lock — see §10. Returns are over the canonical historical window (~6 years for the BTC corpus, shorter where pin years span less).

3 · Why deployed sizing must change

The deployed values you see in /opt/btc_alpha_portfolio/config/strategies/*.yaml and /opt/eth_alpha_portfolio/config/strategies/*.yaml were produced by an original sizing investigation that combined per-strategy half-Kelly with confidence tiering, multi-mode allocation, Monte Carlo stress, and selectivity p-values. That investigation was rigorous — but its inputs were corrupt.

The Trust-Forward Phase 5 work re-pinned every strategy's trade list under correct gate-active conditions and frozen those pins as canonical. Wave 2 reads those pins. The right-hand "Deployed" columns below are what's in YAML right now; the "Proposed" columns are what falls out of the canonical inputs.

The pattern in every per-strategy table

Across all three modes, deployed > proposed for nearly every strategy. The biggest gaps fall on strategies whose canonical Kelly is zero or near-zero — i.e. strategies that don't have a measurable edge once their gate runs correctly. Three are dropped in every mode (b4_fragbreak, s5_smcbias, b1_vol_compression_btc), and three more are dropped in Conservative (b1_vol_compression_btc already at 5%, t1_trend_pullback). One strategy — f1_follow_short — is sized above its deployed level only in Aggressive (250% vs 133%), driven by 108 trades clearing the tier-1 confidence threshold.

4 · Per-strategy Kelly inputs

These are the half-Kelly raw and tier-weighted values. Tier weights: 1.0 (≥80 trades), 0.7 (30–79), 0.4 (10–29), 0.2 (<10). All Kelly outputs hard-capped at 25% per strategy.

Strategyn½-Kelly rawTier-weightedDeployed%
b2_dipbuyer_btc290.07720.0309300%
sma4_btc_gated290.18800.0752260%
b4_fragbreak150.00000.0000253%
s1_gated1230.21600.2160195%
t1_trend_pullback390.01450.0102156%
a3_btc_lead_lag_gated310.08990.0629144%
c5_oct_nov140.30710.1000134%
f1_follow_short1080.26250.2500133%
sma4_eth_gated70.30910.0500125%
continuation_rider250.13490.0540108%
b2_dipbuyer_eth180.25580.100091%
b1_vol_compression_eth610.12290.086083%
s4_momentum_quintile160.11860.047574%
s5_smcbias840.00000.000025%
b1_vol_compression_btc580.00010.00015%

Read this as: s5_smcbias has 84 trades (tier 1.0 weight) but Kelly = 0 — the strategy traded a lot but didn't make money. sma4_eth_gated has only 7 trades (tier 0.2, hardly statistically significant), which is why its raw 0.31 collapses to 0.05 tier-weighted.

5 · Conservative — total notional 107%

✅ Pros

  • Smallest gap to current deployment for the floor of risk — easiest psychological cutover.
  • Drawdown floor: base sim max DD -0.6%, MC P10 still +34.9%. Hard to lose money in this configuration even if every strategy underperforms its canonical sample.
  • Drops 4 strategies entirely — buys time to investigate b4_fragbreak, s5_smcbias, b1_vol_compression_btc, t1_trend_pullback before re-deploying any of them.
  • Frees cross-margin capacity — if leverage compresses required margin, total deployed margin shrinks dramatically vs the current 1900%+ aggregate notional. Capacity for the new strategies arriving from ATLAS proposals.

⚠️ Cons

  • Forfeits roughly 6× of expected return vs Balanced. If Balanced is the right Kelly answer, Conservative leaves money on the table.
  • Tiny per-strategy notionals (mostly 5-10%) mean per-trade P&L is small — a winning month barely registers vs costs. Edge dilution risk.
  • Doesn't actually solve the question — if you're going to halve everything you might as well wait for Wave 2 v2 (selectivity p-values + walk-forward) before committing.
  • Only 11 active strategies — concentrates exposure into the biggest survivors.

Per-strategy breakdown

StrategyProposed%Deployed%Δ
a3_btc_lead_lag_gated6.3%144%↓↓ -96%
b1_vol_compression_btc0.0%5%✗ drop
b1_vol_compression_eth8.6%83%↓↓ -90%
b2_dipbuyer_btc3.1%300%↓↓ -99%
b2_dipbuyer_eth10.0%91%↓↓ -89%
b4_fragbreak0.0%253%✗ drop
c5_oct_nov10.0%134%↓↓ -93%
continuation_rider5.4%108%↓↓ -95%
f1_follow_short25.0%133%↓↓ -81%
s1_gated21.6%195%↓↓ -89%
s4_momentum_quintile4.7%74%↓↓ -94%
s5_smcbias0.0%25%✗ drop
sma4_btc_gated7.5%260%↓↓ -97%
sma4_eth_gated5.0%125%↓↓ -96%
t1_trend_pullback0.0%156%✗ drop

Pick this if: you don't fully trust the canonical pins yet, or you want to bridge to Wave 2 v2 with the smallest possible bet on the new picture being correct.

6 · Balanced — total notional 433%

✅ Pros

  • Tracks half-Kelly directly — each strategy's notional is a 4× scalar over its tier-weighted Kelly with the per-strategy 25% cap respected. This is the textbook risk-of-ruin frontier.
  • Highest expected Sharpe of the three. Aggressive earns more in absolute terms but Balanced earns more per unit of notional and per unit of MC P10/P90 spread.
  • 12 active strategies — diversification floor maintained without forcing capital into Kelly-zero strategies.
  • Drawdown stays below 3% in base sim. Even adding back temporal correlation the realised DD stays bounded if the correlation between strategies is <0.6.
  • Easiest to defend if a third party reviews the methodology — it's "what half-Kelly says, scaled to a sensible aggregate."

⚠️ Cons

  • Still ~75% reduction on most strategies vs deployed — visible regime shift in P&L magnitudes the day you cut over.
  • Doesn't capitalise on the strongest canonical edgef1_follow_short (Sharpe 0.53, PF 8.78, 108 trades, tier 1.0) sits at 100% notional rather than its uncapped allocation.
  • Aggregate 433% notional still requires meaningful margin under cross-margin if leverage is bounded by per-trade caps. Verify the deployment headroom.
  • Inherits the Wave 2 v2 caveats just like the other modes — temporal correlation across strategies and selectivity p-values are not yet modelled.

Per-strategy breakdown

StrategyProposed%Deployed%Δ
a3_btc_lead_lag_gated25.2%144%↓↓ -83%
b1_vol_compression_btc0.0%5%✗ drop
b1_vol_compression_eth34.4%83%↓↓ -59%
b2_dipbuyer_btc12.4%300%↓↓ -96%
b2_dipbuyer_eth40.0%91%↓↓ -56%
b4_fragbreak0.0%253%✗ drop
c5_oct_nov40.0%134%↓↓ -70%
continuation_rider21.6%108%↓↓ -80%
f1_follow_short100.0%133%↓ -25%
s1_gated86.4%195%↓↓ -56%
s4_momentum_quintile19.0%74%↓↓ -74%
s5_smcbias0.0%25%✗ drop
sma4_btc_gated30.1%260%↓↓ -88%
sma4_eth_gated20.0%125%↓↓ -84%
t1_trend_pullback4.1%156%↓↓ -97%

Pick this if: you want the canonical-derived sizing that's defensible mathematically — half-Kelly with confidence tiering, no extra leverage assumption, drops the strategies whose gate-active edge is zero.

7 · Aggressive — total notional 1083%

✅ Pros

  • Captures the full canonical edge. The strongest tier-1 strategies (f1_follow_short, s1_gated) get sized to their uncapped Kelly: 250% and 216% respectively.
  • Closest to current aggregate notional (1083% vs the deployed ~1900%+) so margin and execution behaviour change least.
  • 13 active strategies — broadest active surface area in any mode.
  • Base terminal +1481% over the canonical window. Even MC P10 returns +961% — the tail risk in this mode is "above-target win, not loss."

⚠️ Cons

  • Concentration risk on f1_follow_short at 250% notional. The strategy's pin includes one outlier trade (+453% in 6 months over the 2020-09 to 2021-03 BTC bull leg). Without that single trade, the canonical Kelly drops materially. Aggressive is a partial bet on that outlier being repeatable.
  • Wave 2's MC perturbation does not stress single-trade outliers — it perturbs each return by ±30%, but the +453% trade still anchors the distribution. A truly punishing stress would resample with replacement and drop the top 5% of trades; that's a Wave 2 v2 add.
  • Higher cross-strategy correlation surfaces — Aggressive activates 13 strategies at meaningful sizes simultaneously. If two crypto-long strategies fire at the same time, the realised drawdown will exceed the simulated -5.3% because the shuffled-leg approximation doesn't model that overlap.
  • If the canonical pins themselves are wrong — e.g. if the EMA-22 gate is over-suppressing in some regime — Aggressive fails the loudest.

Per-strategy breakdown

StrategyProposed%Deployed%Δ
a3_btc_lead_lag_gated62.9%144%↓↓ -56%
b1_vol_compression_btc0.1%5%↓↓ -98%
b1_vol_compression_eth86.0%83%≈ +4%
b2_dipbuyer_btc30.9%300%↓↓ -90%
b2_dipbuyer_eth100.0%91%≈ +10%
b4_fragbreak0.0%253%✗ drop
c5_oct_nov100.0%134%↓ -25%
continuation_rider54.0%108%↓↓ -50%
f1_follow_short250.0%133%↑↑ +88%
s1_gated216.0%195%↑ +11%
s4_momentum_quintile47.5%74%↓ -36%
s5_smcbias0.0%25%✗ drop
sma4_btc_gated75.2%260%↓↓ -71%
sma4_eth_gated50.0%125%↓↓ -60%
t1_trend_pullback10.2%156%↓↓ -93%

Pick this if: you trust the canonical pins, you accept that the tail risk in this configuration is upside concentration on one or two strategies, and you want to maximise the canonical edge before Wave 2 v2 refines the picture.

8 · Strategy-level need-to-knows

f1_follow_short — the +56534% question

The pin meta.yaml shows 56534% total return. This is a display artefact, not a realisable production number. The heritage harness assumes 100% notional per trade with full reinvestment across all 108 trades; one outlier (long 2020-09-28 at $10,879 → 2021-03-15 at $60,190, +453%) drives a large fraction of the compound. Wave 2 reads win_rate / avg_winner / avg_loser directly from the trade list and applies half-Kelly — the +56534% never enters the maths. Aggressive's 250% allocation comes entirely from the per-trade Kelly (PF 8.78, WR 59.3%, n=108 → tier 1.0).

What this means for the decision: Aggressive's headline number is real, but its concentration risk on this one strategy is also real. If the long-BTC bull-leg dynamic doesn't repeat, this strategy underperforms its sample. The original meta.yaml caveats are explicit on this point.

The three perma-drops

b4_fragbreak — 15 trades, PF 0.97, Sharpe -0.05, Kelly 0. Every mode drops it. 253% deployed is the largest per-strategy size in the current portfolio and the canonical evidence for the strategy is straight-up negative. Highest priority for "why was this 253%" review.

s5_smcbias — 84 trades (sample is fine), WR 28.6%, Kelly 0. The strategy is statistically reliable in showing it doesn't make money. 25% deployed is small in absolute terms but it should be 0 unless you want to keep it for paper-trading observability.

b1_vol_compression_btc — 58 trades, Kelly 0.0001 (essentially zero). Already nerfed to 5% deployed. ETH counterpart (b1_vol_compression_eth) has Kelly 0.0860 and does survive in all modes — interesting cross-venue divergence worth investigating.

c5_oct_nov and sma4_eth_gated — small-n caps

Both have raw Kelly >0.30 (very strong) but trade counts of 14 and 7 respectively. The confidence tiering caps them at 0.40 and 0.20 weight. This is not a flaw in the canonical pin — it's the tier doing exactly what it's designed to do. Small-sample Kelly is unreliable; the cap is a Bayesian prior asserting "we don't believe seven trades."

If those strategies accumulate more trades over the next quarter and the edge holds, they'll re-tier and earn larger sizings automatically on the next rerun.

s1_gated — the workhorse

123 trades, raw Kelly 0.2160 (already at the 25% cap pre-tier), tier 1.0 weight. The most statistically robust edge in the cohort. In Aggressive it gets 216% — slightly above current deployed (195%). This is the one strategy where the canonical pin and the deployed sizing roughly agree, which is reassuring evidence that the original sizing investigation wasn't completely off — it was off mostly on the gate-broken inputs.

9 · Reading the Monte Carlo result correctly

0.0% liquidation under MC ±30% sounds bulletproof. It isn't. Here's what the MC actually does, and what it doesn't:

✅ What 0% liquidation does mean

  • For each strategy, perturb every per-trade return by a random ±30% factor. Run that perturbed sequence through the portfolio. Repeat 200 times.
  • Across all 200 trials, no run drove cumulative equity below the 5% liquidation threshold from $10K starting equity.
  • This is a meaningful safety check on per-trade noise — outliers don't drag a single strategy into a death spiral.

❌ What 0% liquidation does NOT mean

  • Sample paths matter. 200 trials is a lot for the median but thin in the tails. A 0% rate from 200 trials only tells you the true rate is plausibly <1.5% with 95% confidence — not zero.
  • ±30% is a soft stress. Realistic regime shifts (e.g. a 2018 or Q1 2022 crypto winter) compress win rates and inflate avg loss together — not just multiply each return independently.
  • Temporal correlation across strategies is not modelled. If sma4_btc_gated and a3_btc_lead_lag_gated both fire long during the same 2-week window, the realised drawdown stacks. The shuffled-leg approximation interleaves trades randomly across time and underestimates correlated losses.
  • Single-trade outliers stay in the stress sample. The f1_follow_short +453% trade is preserved (×0.7 to ×1.3) in every MC trial. A more punishing stress would drop it and re-run.

Treat "0% MC liquidation" as a necessary but not sufficient safety check. The decision is still about whether you trust the canonical edge to materialise in production.

10 · Leverage & cross-margin context

The "total notional %" numbers above (107% / 433% / 1083%) are not the same as the margin you'd lock up in production. Crypto Alpha uses leverage in the opposite direction from what most retail traders do — leverage compresses margin, not amplifies exposure:

Practically: 433% aggregate notional under Balanced does not mean you're using 4.33× the account's equity. It means the sum of all simultaneous notional exposure when every strategy is fully active. With 50× leverage on each leg, the actual margin lock at full simultaneous activation would be on the order of 9% of equity. Even Aggressive's 1083% lands around 22% margin lock at 50×.

Important
This makes Aggressive less scary on the margin side than the headline suggests, but it makes the per-trade sizing still very real. A 250% notional f1_follow_short position taking a 3% loss is still a 7.5% account hit even at 50× leverage. Leverage compresses required margin; it does not compress P&L.

11 · What this rerun does NOT model

Wave 2 first-cut limitations — make sure these don't bite later

  • Temporal correlation across strategies. Portfolio simulation uses shuffled-leg interleaving — strategies' trades are randomised in time rather than placed at their actual timestamps. Real correlated drawdowns (multiple strategies long-BTC in the same week) will exceed the simulated max DD.
  • Selectivity p-values not recomputed. The original investigation used p-values to determine which strategies deserved capacity at all. Wave 2 inherits the universe (15 strategies) and skips that filter. A v2 rerun would re-derive p-values on canonical pins.
  • Walk-forward validation deferred. A robust sizing framework holds out a portion of each strategy's trade list and confirms the proposed sizing performs out-of-sample. Wave 2 uses the entire canonical sample to compute Kelly.
  • The 12+15 stress test suite from the original investigation is not run. That suite includes regime-specific stress (bull only, bear only, sideways only) and is what produced the original confidence bands.
  • Funding cost modelling. Per-trade returns in the canonical pins are net of fees but Wave 2's Kelly does not separately model funding-rate variance, which can flip a marginally positive strategy negative in extended sideways markets.

Wave 2 v2 would address these — call it 1-2 days of work for selectivity p-values + walk-forward, longer for the full stress suite.

12 · Recommended path forward

What's actually being decided: do you cut over now to one of the three modes, or do you keep deployed sizing while Wave 2 v2 closes the caveat list?

  1. The deployed sizing is provably wrong in the sense that it was tuned on broken inputs. Continuing to run with it indefinitely is the worst option — you're trading on a configuration nobody has audited against the canonical truth.
  2. Conservative buys time. If you're not ready to commit to the canonical picture, Conservative reduces exposure to a level where regret risk is small, drops the obviously-broken strategies, and preserves margin headroom for Wave 2 v2 + new ATLAS proposals.
  3. Balanced is the technically-defensible answer. If you trust the canonical pins and you're not over-weighting on f1_follow_short outlier risk, Balanced is exactly what half-Kelly + tiering says you should do.
  4. Aggressive is justified only if you're explicitly making a bet that f1_follow_short and s1_gated will continue to fire at their canonical rate. The MC stress doesn't fully cover that bet.
Suggested sequence
  1. Cut over to Conservative or Balanced now — pick whichever matches your conviction in the canonical pins.
  2. Push Stream 8 backport to BTC + ETH heritage remotes (zero runtime impact, future-proofs the canonicalisation pipeline).
  3. Schedule Wave 2 v2: re-derive selectivity p-values, add walk-forward, re-run with timestamp interleaving for true correlation. ~1-2 days.
  4. If v2 confirms Balanced or Aggressive, re-run the cutover with the updated bands.
  5. If v2 dramatically changes the picture (e.g. a strategy you kept turns out to fail walk-forward), you've already de-risked by being below deployed.