Forward paper-trading record, validated strategies, and evaluations of published algorithms. All backtest numbers are walk-forward out-of-sample, after 5 bps taker fees and size-aware slippage. Generated 2026-08-04 09:36 UTC.
Real forward record β daily sessions with simulated fills (fees + slippage) and actual perp funding accrual. No hindsight. The crypto book is one ensemble of three sleeves β trend, cross-sectional momentum/carry, and capitulation reversion β traded together, with a forward-vs-backtest deviation monitor shown on the tile.
| Date | Session | Equity | Since inception | Positions | Fills | Funding | Regime |
|---|---|---|---|---|---|---|---|
| 2026-08-03 | 12 | $94,942.59 | -5.06% | 12 | 13 | $89.43 | bearish |
| 2026-08-02 | 11 | $96,950.29 | -3.05% | 12 | 4 | $127.86 | bearish |
| 2026-08-01 | 10 | $97,258.77 | -2.74% | 12 | 5 | $158.20 | bearish |
| 2026-07-31 | 9 | $97,230.53 | -2.77% | 12 | 16 | $3.11 | bearish |
| 2026-07-30 | 8 | $97,347.10 | -2.65% | 9 | 7 | $7.06 | bearish |
| 2026-07-29 | 7 | $97,944.47 | -2.06% | 9 | 12 | $9.41 | bearish |
| 2026-07-28 | 6 | $102,684.17 | +2.68% | 11 | 11 | $59.33 | bearish |
| 2026-07-26 | 5 | $101,293.80 | +1.29% | 12 | 13 | $50.12 | bearish |
| 2026-07-25 | 4 | $100,386.56 | +0.39% | 9 | 4 | $2.25 | bearish |
| 2026-07-24 | 3 | $100,411.36 | +0.41% | 9 | 9 | $2.63 | bearish |
| 2026-07-23 | 2 | $99,761.70 | -0.24% | 9 | 12 | $32.17 | bearish |
| 2026-07-22 | 1 | $99,936.53 | -0.06% | 11 | 11 | $0.00 | bearish |
Chained out-of-sample equity from rolling walk-forward folds β the training window never sees the test window. Trend Rider: Sharpe +1.28, CAGR +30.4%, max drawdown -18.5% over 61 folds. XSect momentum+carry: Sharpe +1.25, CAGR +33.0%, max drawdown -25.1%. The deployed portfolio blends both (correlation 0.11) at a headline Sharpe of +1.69, max DD -10.8%. That +1.69 is an upper bound: after deflating for the parameter configurations we searched and the ~5-year sample, the honest estimate is β1.2 (plausible range 0.8–1.5) — still well above bitcoin buy-and-hold's 0.55. See Deflated Sharpe ratio in the glossary below.
| Strategy | Data / TF | OOS Sharpe | CAGR | Max DD | Verdict |
|---|---|---|---|---|---|
| Portfolio: Trend Rider + XSect (55/45 inverse-vol) | binance_liquid 1d | +1.69 | +31.0% | -10.8% | Core of live v2 Β· β1.2 deflated |
| Trend Rider (long-only, BTC regime filter) | binance_liquid 1d | +1.26 | +26.7% | -16.3% | Validated Β· transfers to 2 other venues |
| XSect momentum + carry L/S | binance_liquid 1d | +1.25 | +33.0% | -25.1% | Validated Β· survives 2Γ fees |
| Gated funding arb (vol gate) | BTC 1h | +0.40 | +7.0% | -31.8% | Decayed β negative every year since 2023 |
| Grid trading (optimized) | BTC 1h | +0.01 | ~0% | -1.4% | Break-even; fees eat the edge |
| Stat arb BTC-ETH / multi-pair | 1d | -0.12 / -0.58 | -0.4% / -3.1% | -9% / -18% | Cointegration breaks out-of-sample |
| Trend Rider @ 1h (same logic) | binance_liquid 1h | -1.45 | -37.2% | -75.9% | Trend persistence is daily, not intraday |
| Wave Catcher / Elliott / EMA breakout | BTC 1h-5m | -0.8 β¦ -9.8 | β | β | Dead after costs |
| RL (PPO, 2 reward designs, 5 seeds) | BTC 1h | -2.6 median | β | β | High seed variance β overfit |
| ML direction forecasting (XGB/LGBM) | BTC 1h | ~+1.2 pre-cost | β | β | 52-53% accuracy dies to costs |
Known bias: crypto universes contain only currently-listed coins (survivorship). Results are an optimistic upper bound, partially mitigated by long-only bias, volume gates, and replication across three venues.
Strategies from academic papers and replication studies, re-implemented in our event-driven backtester and walk-forward tested on our crypto universes with full costs. Published numbers rarely survive contact with fees, slippage, and out-of-sample discipline β the point of this section is to show which ones do.
| Algorithm & source | Published performance | Our Sharpe | Our CAGR | Our max DD | Verdict |
|---|---|---|---|---|---|
| Time-series momentum, vol-targeted (crypto perps) Moskowitz, Ooi & Pedersen 2012 (JFE); crypto replications 2020-25 30d lookback L/S, vol-targeted, weekly rebalance. Fee-robust (2Γ fees: Sharpe +0.73) but grid-selected, only 48% profitable folds, and same alpha family as our +1.26 Trend Rider. | Sharpe β 1.5-1.7 vol-targeted BTC trend vs 0.8-1.3 buy-and-hold | +0.75 | +27.4% | -53.4% | Works, but redundant with Trend Rider |
| Short-term reversal (crypto alts) Kosc, Sakowski & Εlepaczuk 2019 (Physica A); IRFA 2022 Best long-only variant: -73% max DD and 2Γ fees cut CAGR from +12% to +4.5%. The published L/S construction is Sharpe -0.31 after costs β the premium lives in illiquid names, pre-cost. | Contrarian dominates momentum on 100-coin universes; edge is a liquidity-provision premium | +0.52 | +12.2% | -73.0% | Does not survive costs |
| Funding-rate extreme contrarian (crypto carry) BIS WP 1087 (2023); He et al. 2022 Already implemented here as funding arb β aggregate Sharpe is legacy 2020-22 alpha; negative every year since 2023 as the funding market matured. | Carry Sharpe up to ~6 in 2020-23 samples, decaying to negative by 2025 | +0.40 | +7.0% | -31.8% | Decayed β negative since 2023 |
Plain-language walkthroughs of every algorithm currently trading in the crypto ensemble above β what each one actually does and the reasoning behind it.
How it works. Goes long only on a strong, volume-confirmed breakout: the trailing ~40-day return must be at least +40% AND the day's volume at least 1.5x its 24-day average (price strength backed by a volume surge). Each entry risks 3% of equity to a 3xATR stop, then rides a chandelier trailing stop with scale-outs at the 1R, 2R and 3R milestones, plus time-based and volume-dry-up exits. It stands completely flat whenever BTC is in a bear regime β BTC's close below its 100-day average x 0.95.
Why it should work. Powerful crypto breakouts tend to persist, so it rides the strongest movers, cuts losers quickly, and refuses to fight a bear market. It is deliberately positive-skew: a few big winners carry the whole book, and only about 39% of test windows are profitable.
When it makes money / when it struggles. Makes money in sustained bull trends that throw off a handful of explosive movers; struggles in choppy, sideways or bear markets β where by design it mostly sits in cash.
How it works. Every 7 days it ranks the liquid perpetual universe on a 50/50 blend of momentum (the 30-day return, skipping the most recent 3 days) and carry (favouring low or negative funding β being paid to hold). It buys the top 5 and shorts the bottom 5, inverse-volatility weighted, with about 50% gross exposure on each side so the book's net direction is roughly zero.
Why it should work. Cross-sectional momentum and carry are two of the best-documented, most durable return premia in markets. The market-neutral construction lets the sleeve work in a bull or a bear alike β it kept working through the current bear regime β and it also earns funding. This is the steadiest sleeve, profitable in about 75% of test windows.
When it makes money / when it struggles. Makes money when recent winners keep leading recent losers and funding stays favourable; struggles in sharp reversals where yesterday's laggards suddenly lead, or when the whole universe moves together and dispersion collapses.
How it works. It buys a violent alt crash β a coin down at least 8% over 4 hours on a 2x-or-greater volume spike β but only during a market-wide BTC flush (BTC itself down at least 3% over the same 4 hours). Every trade is bracketed with a 3xATR stop, a 1.5R take-profit and a hard 48-hour maximum hold.
Why it should work. Forced liquidations overshoot to the downside; buying into the flush aims to capture the snap-back bounce once the forced selling exhausts itself. It is a rare-event, insurance-like tail strategy β it sits idle with zero trades most of the time and only fires in genuine market-wide crashes.
When it makes money / when it struggles. Makes money when a panic sell-off overshoots and rebounds within a day or two; struggles when a crash simply keeps going with no bounce β which is why every trade is tightly stopped and time-boxed.
Weights are the crypto ensemble's current target allocation (v2). These are research strategies run in paper trading to study how they behave β described here for transparency, not investment advice.
What we're researching and building next β reported honestly, wins and NO-GOs alike. Finding that a heavily marketed edge is not alpha is a result we keep, not one we bury. Nothing below is wired into live trading.
What it is. Following legally public crypto flows β on-chain whale transactions, perp funding, open-interest flushes, stablecoin liquidity β read faster than the crowd. Public / observable data only; never material non-public information.
Honest verdict. Almost none of it is standalone alpha for a small book. Naive whale-alert copy-trading is front-run by mempool bots and gas / slippage eats the edge; coincident on-chain metrics (SOPR, MVRV-Z, Puell) are redundant and tiny-sample. The Phase-1 backtest of crypto informed-flow signals came back a NO-GO β data-blocked and inside sampling noise.
Where the value is. As free positioning / liquidity inputs to log now β funding z-score, OI-flush confirmation, stablecoin supply, spot-ETF net flow β feeding the capitulation sleeve and a future crypto event dial, not a timing edge.
What we're doing. Forward-recording the free crypto positioning signals above with strict two-timestamp point-in-time discipline (observed_at vs reference_date), joined from the always-on liquidation collector, to self-build history that is unrecoverable later.
Why now. Same logic as the liquidation collector: start the clock even though we don't trade it yet. Cheap optionality; nothing wired to live trading.
What it is. A NEW observability system β not a strategy β that alerts when the live market has drifted away from the regime a strategy was fit to. A leading structural-drift warning that fires before the PnL confirms the damage, complementing the existing lagging monitor (realised equity vs backtest band, shown on the crypto tile above).
The build decision. Four routes were prototyped on real data. Ship the statistical detectors (BOCPD change-point + regime-novelty) and a PCA / Mahalanobis shadow detector; build next the Kalman-innovation monitor β the only route not blind to the trend sleeve's conditional-dependence decay; skip deep nets (no gain over linear on daily data until L2/tick).
Governance. A two-key WATCH → WARN → HALT ladder: a structural alarm must corroborate a performance alarm before any de-risking (input drift alone = early warning; a PnL drop alone = sampling noise). Alarm bands sit outside the ±0.45 Sharpe sampling error, so a 1.1 → 0.6 Sharpe drop β within one standard error β does not trip an alarm.
Honest read. Spans both books (the crypto sleeves and the equity rotation). Risk / observability infrastructure only β shadow-logged, nothing wired to live trading.
Real forward record β daily sessions. The ETF sleeve runs a live head-to-head: our momentum rotation vs GEM dual momentum, the best published algorithm we tested β same universe, same costs, same start date, so the forward difference is attributable to the signal alone.
Two independent ledgers per strategy. The paper sim tiles are our internal fill simulator (orders filled locally against the close with modeled fees + slippage) β a self-contained ledger, not connected to any broker. The Alpaca live paper tile reads the real Alpaca paper account, which only moves when an order actually fills at the broker. They can legitimately differ: the sim always fills, whereas a live order that misses its auction window shows no position change at Alpaca.
Our own cross-sectional momentum applied to US equities, walk-forward tested with full costs. One equity strategy clears the bar; its headline number is an optimistic upper bound until point-in-time index membership removes the survivorship bias β the universe currently uses today's constituents.
| Strategy | Data / TF | OOS Sharpe | CAGR | Max DD | Verdict |
|---|---|---|---|---|---|
| XSect momentum, US equities (long-only 12-1) | us_equities 1d | +0.83 | +20.7% | -53.7% | Survivorship-biased upper bound |
Known bias: the broad US-equity universe uses current index constituents (survivorship-biased) until a point-in-time membership source lands. The fixed ETF rotation basket is not affected.
Strategies from academic papers and replication studies, re-implemented in our event-driven backtester and walk-forward tested on our equity and ETF universes with full costs. Published numbers rarely survive contact with fees, slippage, and out-of-sample discipline β the point of this section is to show which ones do.
| Algorithm & source | Published performance | Our Sharpe | Our CAGR | Our max DD | Verdict |
|---|---|---|---|---|---|
| Faber GTAA (10-month SMA tactical allocation) Faber 2007/2013 (SSRN) Cuts max DD to -16% vs SPY's -51% over the same period, but gives up ~7pp CAGR (cash sleeves modeled at 0% yield β conservative vs the paper). | β10% CAGR with -10-15% max DD, 1973-2012; equity-like returns, bond-like drawdowns | +0.58 | +4.3% | -15.6% | Drawdown control confirmed, no excess return |
| Dual momentum / GEM (Antonacci) Antonacci 2014; extended 1950-2018 backtest {6m,12m} lookback ensemble (single-lookback GEM is timing-luck sensitive). Sharpe +0.73 vs SPY's +0.64 same-period, max DD -25% vs -55%. Now tracked daily in live paper (head-to-head vs our ETF rotation above). | ~15% CAGR, Sharpe ~0.9 in-sample 1974-2013; trails benchmark since ~2012 | +0.73 | +9.3% | -25.0% | Matches SPY with half the drawdown |
| Residual (idiosyncratic) momentum, equities Blitz, Huij & Martens 2011 (JEF); replicated internationally Queued β needs point-in-time index membership to avoid survivorship bias. | Momentum-like returns at ~half the volatility; Sharpe ~0.9-1.0 vs 0.45 standard | pending | Not yet evaluated | ||
Plain-language walkthrough of the equities book currently trading in live paper above β what it does and the reasoning behind it.
How it works. A monthly-cadence cross-sectional rotation across a ~12-ETF multi-asset basket. It ranks each ETF by 6-month momentum (skipping the most recent month to sidestep short-term reversal), holds the top 3 inverse-volatility weighted, and rebalances about every 21 sessions. A SPY 200-day moving-average regime gate flips the book to cash when the broad market is below trend. In live paper it submits market-on-open orders to the Alpaca paper API about 90 minutes before the US open.
Why it should work. Rotate capital toward whichever asset classes are strongest (risk-on) and step aside to cash when the trend is weak. Cross-sectional trend plus a simple regime filter is a well-worn way to capture the up-years while sidestepping the worst of the deep equity drawdowns.
When it makes money / when it struggles. Makes money in persistent risk-on trends where a few asset classes clearly lead; struggles in whipsaw markets that fake out the regime gate, and it gives up some upside by rotating and de-risking a step late.
The US ETF rotation is a separate book from the crypto ensemble. These are research strategies run in paper trading to study how they behave β described here for transparency, not investment advice.
What we're researching and building next β reported honestly, wins and NO-GOs alike. Finding that a heavily marketed edge is not alpha is a result we keep, not one we bury. Nothing below is wired into live trading.
The pivot. Direction changed from a defensive, market-wide macro risk-dial (scale the books' gross down in risk-off) to an offensive, directional, thematic macro / event strategy for stocks & ETFs.
How the v2 strategy works. Detect a macro / commodity / geopolitical shock from free, point-in-time-clean price data → go long the main beneficiary → propagate to the enabler / dependency chain that reprices as a downstream consequence. Canonical example: war → oil ↑ → long energy (XLE) plus defense (ITA) and midstream infrastructure (AMLP), while flagging the losers (airlines, transports) to avoid.
Why the old framing died. Phase 1 proved the risk-dial is genuinely orthogonal to the crypto price gate (a GO for not double-counting) but a NO-GO as a de-risking overlay: on the already market-neutral crypto ensemble it reduces no out-of-sample drawdown and bleeds ~4pp CAGR β it de-risks into the calm periods it cannot protect. A correct answer to the wrong question.
Status. Designed; the universe and dependency map are pre-registered and dated to stop the map being tuned to the curve. The MVP β a free, PIT-clean commodity-shock event study β is not yet built. Detail: macro & event dashboard →
Honest read. This is the hardest-to-systematize idea in the repo β plan against a thin edge or an outright null, not the seductive story. The one defensible kernel is the Cohen-Frazzini economic-link lead-lag (enablers reprice with a days-to-weeks lag). Guardrails before it can graduate: pre-registration, placebo dependency chains, and it must beat both SPY and just buying the headline beneficiary, out-of-sample.
What it is. The equity / political slice of the same public-data program β congressional (STOCK Act) filings, SEC Form-4 insider buys, 13D / 13F, FINRA short interest, and prediction-market odds.
Honest verdict. Naive copy-Pelosi long = the NANC ETF β large-cap-tech beta, not risk-adjusted alpha; 13F cloning is decayed by its 45-day lag; unusual-options and dark-pool trackers are marketing. Our Phase-1 Form-4 / congressional tilt came in at +0.01β0.02 Sharpe, non-monotonic β inside the noise (the insider edge lives in small / microcaps, not the S&P).
What survives. Prediction-market Fed / CPI odds are a free, well-calibrated event-consensus feed for the Lighthouse. One genuine candidate remains untested at scale: insider opportunistic / cluster buys as a small long tilt on the equity book, pending a small / microcap universe.
What we're doing. Forward-recording free equity filing signals β SEC Form-4 insider buys, 13D activist stakes, FINRA short interest β with the same two-timestamp PIT discipline, dated at filing time to avoid look-ahead. Live-smoke-tested; nothing wired to trades.
What it is. A NEW observability system β not a strategy β that alerts when the live market has drifted away from the regime a strategy was fit to. A leading structural-drift warning that fires before the PnL confirms the damage, complementing the existing lagging monitor (realised equity vs backtest band, shown on the crypto tile above).
The build decision. Four routes were prototyped on real data. Ship the statistical detectors (BOCPD change-point + regime-novelty) and a PCA / Mahalanobis shadow detector; build next the Kalman-innovation monitor β the only route not blind to the trend sleeve's conditional-dependence decay; skip deep nets (no gain over linear on daily data until L2/tick).
Governance. A two-key WATCH → WARN → HALT ladder: a structural alarm must corroborate a performance alarm before any de-risking (input drift alone = early warning; a PnL drop alone = sampling noise). Alarm bands sit outside the ±0.45 Sharpe sampling error, so a 1.1 → 0.6 Sharpe drop β within one standard error β does not trip an alarm.
Honest read. Spans both books (the crypto sleeves and the equity rotation). Risk / observability infrastructure only β shadow-logged, nothing wired to live trading.
A plain-language reference for the metrics and jargon used across both tabs, grouped roughly from basic to advanced.