- DIAGNOSIS: The repeated death of our findings under walk-forward is not a methodological error - it is the method telling us that on the axis 'predicting BTC direction over hours/days' there is almost nothing real. Single-asset timing of a liquid major has signal-to-noise near zero. See [[what_doesnt_work]].
- THERAPY = 3 axis switches: (1) Directional -> Structural (get paid for a function instead of guessing); (2) Single-asset -> Cross-sectional (relative instead of absolute); (3) Signal search -> Sizing/Execution (the edge sits in the 'how much', not the 'when').
- BET 1 - Funding carry [SCOUT MEASURED 2026-06-16]: Delta-neutral carry (long spot + short perp) yields net ~13% p.a. on notional / ~6-7% on liquidation-free 2x capital, BTC/ETH/SOL almost identical (+6.7/+6.6/+5.6%) - CROSS-ASSET GATE PASSED, the project's first clean structural edge. BUT the yield is cooling off (2024 22-28%, 2025 5-11%, 2026 YTD BTC +2.5%/SOL -4.1%): a calm carry sleeve, no home run; uncorrelated to directional strategies. Per [[what_works_botty]], the fade-filter variant (b) is already overfit standalone as a raw entry. Cadence bug found: HL funding is HOURLY, not 8h. Detail: [[funding_rate_arbitrage]] . [[funding_rate]].
- BET 2 - Cross-sectional RV [SCOUT MEASURED 2026-06-16]: Long top / short bottom tertile across 20 HL/Binance coins, weekly. Market neutrality CONFIRMED (|BTC beta| 0.038) and effect real (4/5 lookbacks Sharpe +0.40..+0.56, only 7d negative = reversal). BUT fragile & NOT deployable as-is: Sharpe only ~0.4-0.5, maxDD -29 to -61%, 2026 YTD negative across ALL lookbacks (-19..-33%) = classic momentum crash + poor regime. Would need per-leg vol scaling + crash protection before being taken seriously. Detail: [[cross_sectional_momentum]] . [[time_series_momentum]].
- BET 3 - Vol-targeting/regime overlay: NOT a new signal but a sizing layer on top of existing strategies, using our already TIER-1-proven vol tools (vol-forecast 4h IC +0.84, [[vrp]] IC -0.28, vol-targeting Calmar uplift +1.14 on 3/3 strategies - per [[what_works_botty]] NOT yet wired into execution/). Low-hanging fruit.
- BET 4 - Options/positioning [SCOUT MEASURED 2026-06-16]: Honest data reality - the REAL options signals (skew, put/call, smart-money top-trader L/S, OI) are NOT walk-forward-testable TODAY (no free Deribit skew history; L/S+OI only 28 days accumulated, needs ~3 months). Testable on long history: DVOL/VRP (VRP already TIER 1), ETF flows, Fear&Greed, stablecoins. Finding (IC on BTC forward return): the ONLY stable lead = ETF net flows 5d->20d (IC +0.10, annual stability 3/3, decile spread +7.5%) - a genuinely new orthogonal signal, BUT only 2.4y all-bull history, bear test missing. Fear&Greed recently consistently contrarian (IC negative 2022-2026), but regime-dependent. DVOL level/stablecoins unstable. [[vrp]] . [[open_interest]].
- COMMON DENOMINATOR: None of the four bet on 'our BTC forecast being right' - which is why they are structurally different from everything our gates have killed. Three of them (funding filter, vol overlay, options filter) are cheaply testable as an overlay/filter on existing strategies; cross-sectional is the big infra rebuild.
- SCOUT SYNTHESIS (all 3 scouted 2026-06-16, ranked by deployability): #1 FUNDING CARRY - the only one that cleanly passes the cross-asset gate (~6-7% net / safe capital, directionless, tiny DD), best risk-adjusted profile, BUT modest & cooling off -> usable as a calm carry sleeve. #2 ETF-FLOW LEAD (from bet #4) - AUDITED 2026-06-23 (ml/experiments/etf_flow_residual_20d): it does survive the momentum residual test (residual IC +0.126 pooled, not a pure momentum artifact - unlike the 1d case etf_flow_residual) AND naked price momentum does not predict fwd_20 at all -> genuinely orthogonal. BUT the pooled significance is an OVERLAP ILLUSION: 20d windows overlap 95%, effectively only n=30 independent blocks -> non-overlapping IC +0.171 with CI [-0.23,+0.56], p=0.37 = indistinguishable from zero. Verdict INCONCLUSIVE/data-starved: the best of the positioning signals, but 2.4y all-bull + 20d horizon = not validatable, no deploy. Re-audit only with >=1 bear regime + more years. #3 CROSS-SECTIONAL - anomaly real + market-neutral confirmed, but Sharpe ~0.4 + DD up to -61% + 2026 negative -> not deployable without crash protection. Vol-targeting (#3 of the bets) is running in a parallel session. In common: all three are structural/relative, NOT a BTC forecast - the axis is right, even if only funding is immediately usable.
- BUT the honest bar remains: 'very_strong' in the literature does not mean 'large enough net of fees for us'. Before every deploy, compute the signal ceiling (bps/trade x trades/year) and run it through [[walk_forward]] + [[pbo]] + cross-asset gate. 'Robust' = significant, not large.
Why the old approaches feel like a dead end
They are one - but not for the reason you'd suspect.
For months, at its core, we have been doing a single thing: we try to predict the direction of BTC over hours to days by scanning price history for TA patterns. That sounds reasonable, but mathematically it is the hardest quant problem there is:
- BTC is a highly liquid, highly efficient market. Thousands of pros and bots trade it. Every simple, real pattern is arbitraged away immediately.
- The signal-to-noise ratio of a pure directional forecast is near zero. 99% of the move is noise.
- If you search long enough anyway, you always find something in the backtest - but that is almost certainly overfit (fitted to the past, without real predictive value).
And this is exactly where the good news is hidden: our Walk-Forward and PBO gates work. That Divergences: theory, practice and our walk-forward test on BTC died three times, that donchian@2h was killed as a BTC curve-fit, that 5 of 6 ML findings fell apart under walk-forward - that is not a failure of our method. That is our method correctly telling us: on this axis there is almost nothing real.
The answer is therefore not to search harder. It is: switch the axis.
Three axis switches
| From | To | Idea in one sentence |
|---|---|---|
| Directional (guessing where the price goes) | Structural (getting paid for a function) | The edge comes from the counterparty's predicament, not from your forecast. |
| Single-asset (BTC only) | Cross-sectional (many coins relative to one another) | You bet on relative performance, which is far more persistent than absolute direction. |
| Signal search (the perfect when) | Sizing/Execution (the right how much and how) | With tiny edges, position size and execution decide profit/loss. |
The four bets below are concrete implementations of these three axes.
Bet 1 - Funding carry: let the greedy pay you
In plain words. A perpetual future (Perp / Perpetual) is a future with no expiry date. To keep its price from drifting away from the actual spot price, there is a balancing payment every few hours, the funding. The rule is simple:
When more people are long (betting on rising prices), the longs pay the shorts. When more are short, the reverse.
So funding is a direct gauge of greed vs. fear. On Hyperliquid it is settled hourly.
Why this is a real, structural edge. You need no forecast. In a bull market there are constantly too many leveraged longs - they have to pay, no matter what the price does. You collect that payment for holding the unpopular side. That is a risk premium, not gambling.
Measured (scout 2026-06-16, backtesting/funding_carry_scout.py, HL data May 2023-Jun 2026, net of fees, monthly rebalancing):
| Asset | net on notional | net on 2x capital (liquidation-free) | % hours positive |
|---|---|---|---|
| BTC | +13.4 % p.a. | +6.7 % | 87 % |
| ETH | +13.3 % p.a. | +6.6 % | 86 % |
| SOL | +11.2 % p.a. | +5.6 % | 75 % |
Cross-asset gate passed - three assets, almost identical magnitude, all positive -> consistent with a structural (not curve-fit) edge. BUT: the yield is cooling off (gross: 2024 22-28 %, 2025 5-11 %, 2026 YTD BTC +2.5 % / SOL -4.1 %) - as the market matures, the premium compresses. On safe 2x capital ~6-7 % p.a.: a calm, directionless carry sleeve, no home run, but uncorrelated to the directional strategies. Cadence note: HL funding is HOURLY (not 8h, despite the filename _funding_8h) -> annualize with 24x365, not 3x365.
To harvest it directionlessly (delta-neutral): you buy 1 BTC in spot and simultaneously short 1 BTC in the perp. If BTC rises, the spot gains what the short loses - the price move cancels out, and you keep only the funding. That is the classic Funding Rate Arbitrage (Delta-Neutral) (closely related to the Cash-and-Carry / Basis Trade).
What this means for Botty - two variants: - (a) Pure carry harvest (delta-neutral): needs a spot leg + perp leg and rebalancing logic. More execution, but 'set-and-forget', no signal intelligence required. - (b) Funding as a filter (cheapest to test!): extremely positive funding = crowded longs = elevated reversal risk. We attach this as a contra-filter to our existing directional strategies - no new execution leg needed. Exactly the 'funding rate filter' gap from our Pifagor rebuild.
Risks. Funding can flip negative (then you pay); with delta-neutral there is basis risk and liquidation of the perp leg under poor margin management; fees on both legs.
Bet 2 - Cross-sectional relative value: flip the question
In plain words. Instead of asking "Is BTC going up?" you ask "Which coins are performing best relative to the others - and which worst?" Then:
Buy the strongest (top decile), short the weakest (bottom decile), equally weighted, so that long and short cancel out in aggregate (market-neutral).
Why this is more robust than BTC timing. This is the momentum anomaly - the best-documented anomaly in finance since 1993, robust across 58 futures markets (Cross-Sectional Momentum, Time-Series Momentum (TSMOM)). The point: relative strength is far more persistent than absolute direction. It is much easier to say "SOL is outperforming DOGE" than "BTC rises tomorrow". And because you are equally long and short, a crypto crash barely matters to you - it hits both sides.
Concrete example. Universe = the top-20 perps on Hyperliquid. Every Monday: 1. Sort them all by their return over the last 30 days. 2. Over 30 days: SOL +40 %, ETH +12 %, BTC +8 %, ... , LTC -15 %, DOGE -22 %. 3. Long the top 5 (SOL, ...), short the bottom 5 (DOGE, ...), BTC/ETH in the middle = neutral. 4. Hold for a week, then re-sort.
You earn when the winners stay winners and the losers stay losers - regardless of whether the overall market rises or falls.
Measured (scout 2026-06-16, backtesting/xsectional_scout.py, 20 coins, 2021-2026, net of costs): The market neutrality works (|BTC beta| on average only 0.038 - truly directionless), and the effect is real (4/5 lookbacks positive, Sharpe +0.40 to +0.56; only the 7-day lookback is negative = short-term reversal, expected). BUT: Sharpe stays meager at ~0.4-0.5, the drawdowns are brutal (-29 % to -61 %), and 2026 YTD is negative across all lookbacks (-19 % to -33 %). That is the notorious momentum crash property (momentum breaks down hard at sharp turns) plus a currently unfavorable regime. Conclusion: real, but not deployable as-is - it would need per-leg vol scaling + momentum-crash protection (e.g. switching off after market vol spikes) before risking real capital.
What this means for Botty. This is the biggest infra pivot of the four - but it recycles our entire megasweep/walk-forward machine. For the cross-asset gate we already have multi-asset data and the runner.BINANCE_SYMBOL monkeypatch; we generalize it from "swap one symbol" to "rank across a universe".
Risks. Crowding (many do it); transaction costs from weekly rebalancing; in crypto, correlations are high during panic phases (everything falls together -> spread compresses).
Bet 3 - Vol-targeting overlay: not when, but how much
In plain words. For months we have been hunting for the perfect entry signal. This bet ignores the signal entirely and turns a different knob: the position size.
Make the position smaller when the market is wild, and larger when it is calm - so that your risk always stays the same (constant target volatility).
Why this works - and why especially for us. A smooth risk curve has a higher Calmar Ratio and Risk/Reward than one that swings from huge to tiny - with no better signal at all. And the best part: we already have the tools and they are TIER-1-proven for us. Per What Demonstrably Works in Botty — Evidence Ranking Across All Sweeps, Tests & ML Experiments:
- Vol-forecast (GBM, ml/forecast/): 4h IC +0.84, cross-asset confirmed.
- VRP: IC -0.28 in 16/16 windows. BOCPD changepoints: IC +0.16 in 21/21 windows, live.
- Vol-targeting: Calmar uplift +1.14 on 3/3 tested strategies - but NOT yet wired into execution/.
We have the proof and the module. We just haven't built it in yet. This is the lowest-hanging fruit of the four.
Concrete example. Target: 15 % annualized vol. The vol model predicts high vol for the next 4 h -> halve the position size. It predicts calm -> full (or larger) size. Over the year this smooths the equity curve and lifts Sharpe/Calmar - with an identical entry signal.
What this means for Botty. An overlay layer between signal and order sizing that wraps our existing strategies. Combinable with the regime gate (strategy fully off in bad regimes). No new alpha signal - a multiplier on what we already have.
Risks. Vol-of-vol and lag (the model trails a jump); but the evidence for it is already harder in our own data than for any entry signal we have ever tested.
Bet 4 - Options/positioning signals: what the pros fear
In plain words. The options market (for crypto above all Deribit) is a huge betting shop on what happens soon. From options prices we read off what pros expect and fear - before it shows up in the spot price. Four readings: - Skew: Are puts (crash hedges) more expensive than calls? -> fear. - Term structure: Does the market expect more movement short-term than long-term? -> an upcoming event/stress. - Put/call ratio: How much is being bet on down vs. up? - Open-interest walls (Open Interest): At which strikes does the most money sit? -> magnet/pin levels.
Why this is a leading indicator. Options are forward-looking - they price in the expected future volatility. The systematic premium of expected over realized volatility is the volatility risk premium (VRP) - already promoted and TIER-1 for us (IC -0.28, 16/16 windows). It is the first building block of this family to have passed the shuffle test for us.
Measured (scout 2026-06-16, backtesting/positioning_ic_scout.py). Honest data reality first: the real options signals (skew, put/call) are not available for free as history, and smart-money positioning (top-trader L/S, OI) has only accumulated 28 days - both not walk-forward-testable today. What has long history, tested on forward-return IC:
| Signal | best IC | annual stability | decile spread |
|---|---|---|---|
| ETF net flow 5d -> 20d | +0.10 | 3/3 | +7.5 % |
| Fear&Greed (contrarian) | -0.1 to -0.7 (2022-26) | regime-dependent | - |
| DVOL level / stablecoins / VRP directional | < 0.1, unstable | weak | - |
The only stable, sizable lead is ETF flows - institutional net demand, a genuinely new signal orthogonal to price TA. But caution: only 2.4 years of history (ETFs since Jan 2024), entirely a bull market -> the bear-regime test is missing, 3/3 years is a low bar.
What this means for Botty. (1) Pursue the ETF-flow lead further as a filter/bias layer, but with the bear caveat. (2) Now run the Binance derivatives accumulation (download_derivatives.py) so that smart-money L/S + OI are walk-forward-ready in ~3 months. (3) Skew/put-call need a (possibly paid) historical options data source before they are testable.
Risks. Data plumbing and history are the real bottleneck here - not the idea. Options data is noisy; the translation from 'positioning signal' to 'BTC spot edge' has to pass cleanly through the gates.
How we test this (so it doesn't become a dead end again)
Every bet runs through the same gates that have kept us from fooling ourselves so far - Walk-Forward over multiple calendar years, PBO, cross-asset generalization, and the signal-ceiling calculation (bps/trade x trades/year, net of fees) before anything sees live capital.
Order by effort (cheap -> expensive): 1. Funding filter (variant b) - a filter on existing strategies, no new leg. Testable today. 2. Vol-targeting overlay - the module exists, just wire it. Biggest leverage per effort. 3. Options/positioning filter - some data plumbing (Deribit), then a filter test. 4. Cross-sectional RV - the big rebuild, but it recycles the megasweep infra; its own sweep.
Honest expectation. 'very_strong in the literature' != 'large enough for us net of fees'. Some will die here too - but they die for different reasons than BTC timing, and even a death delivers real knowledge instead of more curve-fitting. The probability that at least one of these structural/relative axes carries a real, deployable edge is considerably higher than another round of TA parameter search on BTC.