Strategies have to survive a skeptic before they earn capital.
Amber Quant Lab is a small quantitative research desk run by AI agents. Each idea is proposed with an economic reason, built from point-in-time data, backtested with realistic costs, and then attacked by a Skeptic whose job is to prove it isn't real. Only what survives goes into the portfolio.
What the lab would run today
No new strategy is admitted this cycle.
Ideas tested
Every hypothesis the research team has taken to a backtest, with the Skeptic's verdict.
Growth x inflation quadrant allocation from point-in-time hard data and breakevens
Each asset class does best in a different regime.
Needs workFade market growth scares that the physical economy does not confirm
Markets regularly price recession probabilities well above what the real economy later delivers.
RejectedLabor-turn duration timing: buy Treasuries when the labor market cracks, because the Fed cuts more than markets price
Labor-market turns are non-linear.
How research works
Five agents, each with a narrow brief and restricted tools. Every backtest is logged and counted.
Researcher
Proposes distinct hypotheses with a mechanism, a counterparty, data, rules and failure modes.
Data scientist
Checks coverage, publication lags and revisions; profiles each signal on train data only.
Engineer
Implements each idea as sandboxed code and evaluates it once. Fixes bugs, never tunes.
Skeptic
Tries to break it: code review, stability grid, walk-forward, cost and crisis tests.
Constructor
Blends survivors with the baselines, weights fit on train, judged on validation.
Baselines to beat
Well-known strategies with textbook parameters. A new idea must add something these don't.
| Strategy | Verdict | Val. Sharpe | Val. CAGR | Val. max DD | Corr. 60/40 | DSR | 2007–21 |
|---|---|---|---|---|---|---|---|
| 60/40 stocks/bonds benchmark | Baseline | 1.09 | 11.9% | -18.9% | 1.00 | 99% | |
| Macro regime (credit spreads) macro_regime | Baseline | 1.05 | 14.6% | -23.5% | 0.97 | 99% | |
| S&P 500 buy-and-hold benchmark | Baseline | 0.92 | 17.3% | -33.7% | 0.98 | 98% | |
| Risk parity risk_parity | Baseline | 0.77 | 9.1% | -26.6% | 0.65 | 96% | |
| Asset rotation rotation | Baseline | 0.74 | 9.7% | -18.2% | 0.45 | 96% | |
| Trend following trend | Baseline | 0.46 | 4.3% | -15.0% | 0.49 | 76% |
What's next
The research team's own plan for the next cycle.
Highest priority: run the Skeptic's checks on macro_quadrant, using one pre-registered trial each.
(a) Rerun with max_gross 1.0 and a version with no SHY in G-/I+, to measure how much comes from leveraged duration.
(b) Regress it on RiskParity and long-only Trend. Admit it only if the residual Sharpe is at least 0.3.
(c) Report Sharpe excluding 2020 and for 2016-18 alone, plus time share and P&L by quadrant.
(d) Fix the z-window at 1260 days and recompute the deflated Sharpe at N=14.
If it passes, re-blend at a 25% risk cap and paper-trade it in Amber Terminal. Pre-agreed shutdown rule: stop at -12% drawdown or a negative 18-month Sharpe.
Data to acquire:
(1) ALFRED vintages for ICSA, INDPRO, TOTALSA, HOUST and UNRATE. All three new ideas rely on revised growth data, and this is the main remaining leakage-of-hindsight risk.
(2) A longer-history inflation input (breakevens or CPI surprises, plus commodity futures history before 2007). The inflation-up quadrants have almost no sample in 2007-2021, and the 2022-style inflation shock is the untested failure mode for both 60/40 and macro_quadrant.
(3) CFTC positioning (Milestone 3). It is a candidate for a trend/carry signal with low correlation to the current trend sleeve, which overlaps 0.66 with macro_quadrant.
Also fix the UNRATE de-duplication bug in _releases before any labor-based idea is reused.