Data refresh in 0:10
Strategies (19)
Pause/start take effect on the next engine restart. Crypto lab strategies link to a detail page with adjustable parameters. See the costs page for the fee hurdles every strategy must clear.
EMA(9/21) crossover, but only taken when price is above rolling VWAP and signed trade flow confirms. Deliberately a reproducible control benchmark, not a claimed edge — richer strategies must beat it net of cost.
- + Very fast to implement and debug; few features, easy to inspect for leakage.
- + Provides a stable control for A/B tests.
- + Works from 1-minute bars if order-book work is delayed.
- − Crossovers lag and overtrade in chop.
- − Common parameters are heavily data-mined.
- − Adds little unique information beyond recent return; can look profitable only from bar-close fill assumptions.
Trades a genuine expansion out of a compressed range only when signed trade flow, microprice displacement and spread state all confirm the move. Long-only spot; the transparent first baseline that establishes the shared execution pipeline.
- + Clear causal story and few parameters.
- + Useful trade frequency without sub-millisecond latency.
- + Works as long-only spot; momentum evidence exists (though regime dependent).
- − False breaks cluster in noisy markets.
- − Taker entry can destroy the edge on Alpaca — targets must clear the fee floor.
- − News and liquidation jumps change the fill distribution.
Enters passively with a resting limit at the bid to earn the maker fee rather than pay the taker fee, in a benign uptrend where flow is not against us. Every fill is protected by a resting stop; unfilled entries are abandoned after a short TTL. The execution-edge candidate — the direct A/B against the incumbent scalp.
- + Earns the maker side of the ~20bp round-trip fee gap — often larger than the whole scalp edge on Alpaca.
- + The cost gate clears a much lower hurdle than a taker entry, so more setups are economically viable.
- + Passive entry avoids paying the spread, and only fills that actually protect are ever held.
- − Adverse selection: passive fills cluster right before the price moves against you.
- − Missed fills — a resting order in a trending market may never fill, so trade frequency drops.
- − Shadow fill modelling of a maker entry is approximate; the real fill distribution only shows in the live paper slot.
Fades a downward liquidity flush — price poking the rolling low on strong aggressive selling, then the book refilling on the bid — targeting reversion toward VWAP with a stop just below the sweep low. With the Phase 5 order-book feed it confirms the sweep from L2 depth imbalance; without a book it falls back to the degraded microprice-rebound proxy (a conservative lower bound).
- + Reversion after a stop-run is one of the more robust intraday crypto edges.
- + Targets a move back to VWAP — usually large enough to clear even taker costs.
- + With book depth it separates a genuine sweep (bid replenishing) from a real breakdown — the main degraded failure mode.
- − Alpaca depth is coarse and can lag; book confirmation reduces but does not remove false fades in a true breakdown.
- − Taker entry to catch the turn pays the spread exactly when it is widest.
- − Catching a falling knife: a fresh low right after entry is the main failure mode.
RESEARCH MODE — does not trade. Reconstructs the L2 order book and records (OFI z-score → forward return) observations for offline analysis. Graduates to a trading candidate only if forward return rises monotonically across OFI buckets net of modelled Alpaca cost (core assessOfiMonotonicity). Requires the orderbooks channel (CRYPTO_ORDERBOOK_ENABLED).
- + Order-flow imbalance is one of the best-documented short-horizon predictors of price on venues with real depth.
- + Signal-only means it can be validated safely with zero execution risk before a single dollar is committed.
- + Produces a falsifiable, bucketed forward-return curve — a clear go/no-go gate rather than a vibe.
- − Alpaca gives no queue position and a coarse book, so the signal is a weaker proxy than the playbook assumes.
- − Taker-only crypto fees will likely swamp the raw edge even if the signal is real — hence research-only here.
- − Book snapshots can lag or cross; observations are only recorded on a valid, uncrossed book.
Round 2 Strategy designs
The post-scalp redesign. S9/S10 run in the crypto lab now; the equities and research designs are fully specified and shown here but await their data path before they can be fed data.
Multi-hour trend-continuation swing on the high-movement alts (AVAX/ADA/SOL/ETH). Enters pullbacks toward VWAP inside an EMA uptrend with flow confirmation, targeting a 2–3% move that clears Alpaca cost ~3× — the direct answer to the Round-1 finding that sub-1% scalps cannot pay their own fees.
- + Attacks the term Round 1 proved broken: a 2–3% target dwarfs the ~0.8% fixed cost, unlike a 0.06–0.37% scalp.
- + Concentrates on the alts that actually move (3–6× BTC), and skips low-vol names via the vol floor.
- + Continuation on pullbacks, not breakout chasing — the Round-1 breakout gate never once fired (0/606).
- − New overnight/gap exposure: a multi-hour hold rides regime changes a scalp never saw.
- − Fewer, larger trades ⇒ slower statistical convergence; needs a longer soak to judge.
- − Trend-continuation buys strength, which mean-reverts in choppy regimes — the vol floor only partly mitigates this.
Fades a capitulation extreme on a high-beta alt — a fresh rolling low on heavy signed selling where the order book is starting to firm — and holds 15–60 min for the bounce. The one Round-2 design the Round-1 data positively hinted at (momentum decayed into reversion). Note: needs a low-fee venue to be broadly viable; on Alpaca it correctly stands down except on rare large-bounce setups.
- + The only design supported by a positive signal in Round-1 data (reversion after short-horizon momentum; faintly contrarian OFI).
- + Buys forced liquidation, not strength — the opposite of the incumbent that went 0-for-154 buying momentum.
- + Book-firming gate avoids catching a one-sided collapse mid-flush.
- − Fading a downtrend is structurally dangerous — a real trend change turns a fade into a full stop-out.
- − Uneconomic on Alpaca fees at realistic reversion sizes; genuinely needs a cheaper venue.
- − Requires a valid order book (depth channel); degrades to no-trade without it.
Multi-day trend/momentum swing on liquid large-cap ETFs and megacaps, entered on a fast/slow MA cross inside a risk-on regime with ATR-scaled stops/targets. Alpaca equities are $0 commission, so the ~0.2% hurdle is trivially cleared by a routine 1–2% move. Runs on daily bars via the equity lab data path; long-only.
- + Removes the fee term entirely ($0 commission) — the single biggest structural escape from the Round-1 cost problem.
- + Reuses infrastructure already built and paid for; free IEX daily bars are adequate for the signal.
- + PDT (<$25k) and T+1 settlement both force the multi-day horizon that ALSO fixes the economics — two constraints agreeing.
- − Introduces overnight/weekend gap risk this system has never carried.
- − Daily-bar cadence ⇒ very low trade frequency; judging it needs weeks-to-months of live bars, not a short session.
- − Free IEX data is fine for daily-bar swing signals but not for anything microstructure; the shadow book only sees daily closes, so intraday stop/target touches are approximated.
Multi-day swing driven by bounded, typed signals from the news/LLM research pipeline — an information edge at a slow horizon, the one domain where a retail system is not outgunned on microstructure. Every order still passes the full risk engine (research is advisory). AWAITING DATA PATH: needs research signals wired into the lab context; specified and registered, not yet run.
- + A genuinely different edge source (information, not microstructure) — where Round-1 proved we are structurally disadvantaged.
- + Slow horizon makes latency irrelevant and the fixed cost trivial.
- + Reuses the research package the PRD already specifies; falsifiable against a random-time-matched control.
- − Needs research signals plumbed into a lab strategy context — no such accessor exists yet.
- − News edges decay fast and are crowded; strength calibration and freshness are critical.
- − LLM/news signals can be noisy or wrong; the advisory-only + risk-engine guardrails are essential.
Round 3 Strategy designs
The multi-day trend follower. The horizon×cost feasibility screen showed the 2-min scalp is dead a priori but Alpaca crypto is viable at a ~1-day hold, where the median move already clears the fixed cost — this trades that directly.
Multi-day (24–72h) trend-following hold on crypto spot, targeting a 5–10% move that dwarfs Alpaca’s fixed ~0.75% round-trip cost. Enters an established daily-scale uptrend (price above a ~1-day EMA) confirmed by multi-hour momentum, but NOT overextended above trend. Directly trades the feasibility-screen finding that the horizon — not the venue — was the wall; the longer-horizon successor to alt-swing.
- + Trades the feasibility-screen result head-on: at a 1-day hold the median move already clears the taker hurdle for every symbol (66–81% oracle clear-rate).
- + Drops the tick-level flow gate that Round 2 proved is noise and was the wall on every directional arm — trend regime + slow momentum instead.
- + A 5–10% target makes the fixed ~0.75% cost negligible (~10× margin), so cost is no longer the binding constraint.
- − Multi-day holds carry real overnight/weekend gap and regime-change exposure a scalp never saw — the wide stop is the only protection between rebalances.
- − Very few, very large trades ⇒ slow statistical convergence; judging this needs a multi-week soak, not an overnight one.
- − Trend-following buys strength, which whipsaws in range-bound regimes; the extension cap and momentum gate only partly mitigate this.
- − Long warmup: the ~1-day trend EMA + 12h momentum lookback mean a symbol records no entries until it has ~1 day of bars.