Slipstream is a quality, trend-adaptive equity paper fund that drafts the smart money — it holds the large-caps that concentrated, high-conviction 13F managers are collectively buying, keeps only the most profitable companies in each sector, caps any one stock at 10%, and rotates into momentum when the market trend breaks. What follows is the rule, the backtest, and an honest account of everything we tried on the way here.
1 · Strategy & approach
The core question: does following the smart money actually work? Our October 2026 re-validation, run from a fresh data pull, found that the original rule (curated super-investors + insiders, ROE/debt quality gate) had no demonstrated edge over SPY. Eight pre-registered follow-up studies then identified what does carry information: breadth of conviction buying among concentrated managers, filtered for profitability within each sector. The rule (v2, live since 5 October 2026), applied monthly:
| Managers | Every 13F filer that behaves like a concentrated stock-picker that quarter: 10–150 positions, $250M–$100B reported equity, ≥40% in its top 10, and a filing the previous quarter. Rule-based, ~900–1,500 managers a quarter — new managers join automatically, nobody curates the list |
| Signal | Consensus: for each stock, 2 × managers opening a conviction position + managers adding − 2 × managers exiting (a conviction position is ≥$1M and ≥0.5% of the manager's portfolio). Buy candidates have a positive count. 13F only — insider buying is no longer used |
| Universe | Top ~1,000 US stocks by market cap, point-in-time |
| Quality | Return on assets in the top 30% of its own sector (latest filing public at the time) — so every sector contributes its most profitable companies, and banks or utilities aren't screened out wholesale |
| Construction | Top 20 by consensus, weighted by the dollars managers committed, each stock 1–10%, at most 5 per sector |
| AI event-risk overlay | A second-stage AI screen reads recent news on the final candidates and hard-excludes only critical events — threats to the company itself: going-concern / bankruptcy, fraud or restatement, criminal indictment, loss of the license to operate, sanctions cutting off a major share of revenue. Lesser events are shown and held. Excluded names are replaced by the next-ranked candidate |
| Dynamic sector cap | Each sector capped at 30% ± its 12-1m relative strength vs SPY (band 15–45%) |
| Regime overlay | When SPY is in a confirmed downtrend (below its 200-day MA, confirmed a month earlier), the whole book rotates into the 50 large-caps with the strongest 12-month momentum |
| Rebalance | Monthly (~3.5× one-way turnover a year), 1% cash buffer, 0.5% trade deadband |
Why these pieces: counting how many conviction managers are buying works better than following any curated few — no 13F manager's individual record proved persistent enough to pick on. The within-sector ROA filter is the single biggest improvement we found; it lifted every manager list we tested. The 10% stock cap keeps the same risk-adjusted return as 15% with a less concentrated book. The regime overlay is what carries the fund through bear markets.
On the AI overlay — deliberately narrow. Selection stays 100% systematic; the LLM is not a stock-picker. Because no backtest can include it, it now fires only on critical events, so the live fund deviates from the tested rule only when a company itself is under threat. It is capped (≤3 exclusions/rebalance), every exclusion cites its headline, and if the news feed or model is unavailable the fund trades the pure quant book unchanged.
2 · Backtest results
Survivorship-free backtest, 2014-03 → 2026-10, rebuilt from a fresh Sharadar pull (4,323 delisted names included; delisting returns from corporate actions). Trades use only data available the day before, net of 20 bp per side, 1% cash; Sharpe and Sortino are net of T-bills. The AI overlay is not included (it can't be backtested).
| Period | Fund CAGR | SPY | Excess | Sharpe (SPY) |
|---|---|---|---|---|
| 2014–2019 (calm bull) | 13.1% | 12.2% | +0.9% | 0.77 (0.88) |
| 2020–2022 (COVID + rate shock) | 11.5% | 7.6% | +3.9% | 0.49 (0.39) |
| 2023–2026 (AI mega-cap rally) | 39.1% | 22.1% | +17.0% | 1.46 (1.11) |
How the rule was chosen: settings were selected on 2016-03 → 2021-12 only (Sharpe 1.08 vs SPY 1.00). 2022 onward was then reviewed across several studies, so it is not a clean out-of-sample test (Sharpe 0.86 vs 0.52). The two years before selection, 2014-03 → 2016-02, were never used — and the fund trailed SPY there (−1.8% vs +4.4% a year, hurt by risk-off whipsaws in 2015–16).
Read honestly. This is a simulated backtest, not a live track record, and it is the best of ~210 configurations tested — so expect live results below it. Over the full period the alpha is positive but only borderline significant (t 1.8); the fund carries more market risk than SPY (β 1.17, volatility ~22%, drawdowns to -37.5%); roughly half the book is usually tech-related; and much of the excess return came after 2023. Candidate rules are tracked in a forward "shadow book" whose books are recorded before their returns exist — that record, not this chart, is the real test. Not investment advice.
3 · Everything we explored
We stress-tested the idea from every angle before settling on the rule above. Each row is a real backtest; the pattern that emerged: every "buy it cheap / beaten-down" tilt hurt, and every "quality / large-cap / diversify / adapt" control helped.
| What we tried | Result | Verdict |
|---|---|---|
| Raw 13F + insider, whole market | 0 of 120 constructions beat SPY; insider buying actively destructive (−63% drawdown, worse than random) | ✗ no edge |
| Prior IC studies (cross-sectional) | Insider IC ≈ −0.02, guru ≈ +0.02 (not significant); only institutional breadth marginally positive ("size proxy") | ✗ confirms weak |
| Restrict to large caps (top-500 / 1,000) | An edge appears (+2–3% excess) — but it was concentrated sector risk, not skill | ~ misleading |
| Sector cap / sector-neutral | The "edge" collapses to ≈ market once the sector bet is removed → proves it wasn't stock-picking | ✗ revealing |
| Dynamic sector cap (relative-strength) | Beats a hard cap: lets leaders run, trims faders; full-period Sharpe 0.82 → 0.86 | ✓ adopted |
| Dip / value entry (below 52w-high or 200DMA) | Made it worse (18 → 1 SPY-beaters) — the falling-knife effect; buying strength beats buying dips here | ✗ rejected |
| Quality screen (ROE, debt/equity) | The least-bad overlay — quality is a real factor; nudged more configs above SPY | ✓ adopted |
| Hold-until-sell (mirror guru holding periods) | Sell signals are rare → book ballooned to ~190 names and diluted to market; underperformed | ✗ rejected |
| Monthly vs quarterly churn | Monthly ≈ quarterly return but more turnover — 13F updates quarterly, so churning adds cost, not signal | ~ no gain |
| Momentum factor (in the 2020–22 window) | The one thing that beat the fund and SPY in that stretch — because it's regime-adaptive | ✓ informed overlay |
| Regime switch → momentum (200DMA, confirmed) | Fixes 2020–22 (−2.8% → +6.0%), lifts full-period excess to +5.0%, same drawdown | ✓ adopted |
| Multi-signal regime consensus (5 signals, K-of-N) | No better than the simple single 200DMA filter; more knobs = more overfitting risk | ~ kept it simple |
| Dynamic cap + regime overlay (v1, Aug 2026) | Reported 19.5% CAGR / Sharpe 0.92 — not reproducible (its 13F input was overwritten); rebuilt on fresh data with live execution: 15.5% CAGR, Sharpe 0.67 vs SPY 0.73, no significant alpha | ~ superseded |
| Wider / different manager lists (102, 210 curated; top-200 track records; "emerging" managers; top inflows) | No list beat the simple rule-based concentrated-manager list; past 13F skill doesn't persist (rank correlation ≈ 0 across 44 quarters) | ✗ rejected |
| Accumulation Radar (volume footprint nowcast of 13F buying) | Detects institutional accumulation well (AUC 0.76) but adds no return on top of 13F consensus | ✗ rejected |
| Quality = ROA, top 30% within sector | Lifted every manager list (e.g. Sharpe 0.89 → 1.07 on 2016–21); smart-money book now beats SPY in risk-on months | ✓ adopted |
| Add debt/equity screen | Lowered return in all 4 variants, no drawdown benefit — mostly removes high-ROA buyback compounders | ✗ rejected |
| 10% stock cap (vs 15%) | Same Sharpe, top-5 weight 56% → 48% | ✓ adopted |
| Tech cap (Technology + Communication ≤ 30–40%) | Costs 1.4–2.5%/yr, no drawdown benefit — a view on tech, not a free risk reduction | ~ optional, not used |
| Concentrated-manager consensus + ROA-30 + 10% cap (v2) | 19.9% CAGR vs SPY 13.9%, Sharpe 0.85 vs 0.73, -37.5% max DD (2014–2026) | ★ Slipstream |
A recurring lesson worth keeping: the flashiest backtest numbers came from the configs taking the most uncompensated risk (concentration, single regime). Each time we tightened the methodology — survivorship, costs, sub-period robustness, sector control, multiple-testing thresholds — the "edge" shrank toward the market. Slipstream is what was left standing after all of that.
4 · Data & honesty notes
- Data: SEC 13F (Sharadar SF3, every filer), fundamentals (SF1), prices (SEP), corporate actions (delistings) via Nasdaq Data Link. Survivorship-free. Sector labels are current, not historical. Form 4 insider data (SF2) feeds the Insider picks panel only.
- Signal freshness: 13F is quarterly and ~45 days lagged by law — so the smart-money signal is inherently stale; the fund is built to work despite that, not to pretend otherwise.
- Picks panels are informational and LLM-screened for manipulation/pump patterns; they are not individual buy recommendations. The fund trades a disciplined, risk-controlled subset automatically.
- Paper only. Slipstream runs on an Alpaca paper account. Nothing here is investment advice.
5 · Operations & monitoring
The fund runs fully automated in the cloud, on Hugging Face scheduled Jobs, with a flag-only supervisor watching it daily. Three independent scheduled loops:
| Weekly refresh (Sun) | A private worker Space re-pulls all source data from scratch (13F, insider, fundamentals, prices, macro), rebuilds the signal panel, and a freshness gate asserts every dataset actually advanced — a stale input fails the run loudly instead of silently trading on old data. |
| Monthly rebalance (1st) | Trades the freshest weekly signal row through the construction rules above (ROA quality gate → AI critical-event screen → 10% name cap alternating with dynamic sector caps). Idempotent — a re-run trades only target-minus-actual. |
| Daily supervisor (every day) | A deterministic health monitor — it never trades — checking three things: tracking (trailing 4-week fund-vs-SPY gap; WATCH <−3pp, ALERT <−6pp), guardrails (paper-only invariant, no margin, exposure ≤100.5%, and position drift — the 10% cap applies at rebalance, so intra-month drift only flags above 13%, alerts above 17%), and freshness (signals ≤10 days old, last rebalance ≤35 days). Results appear as the OK / WATCH / ALERT chips at the top of the Fund tab, and an ALERT opens an alert on the Space so it can't be missed. |
Design choice: monitoring is flag-only. The supervisor observes and reports; only the monthly rebalance ever places orders, and only on the paper account. An autonomous recalibration agent (self-adjusting strategy parameters under strict acceptance gates) was evaluated and deliberately deferred until the fund has a walk-forward variant-testing harness and at least a quarter of live tracking history to calibrate against.