Empirica Score · Phase A complete · 2026-05-26
We built three open factor proxies (Relative Strength, Technical, Fundamental) on free inputs, blended them via regression on 800 paired observations, and validated head-to-head against a publicly-visible third-party scoring benchmark. Two out-of-sample snapshots both came in above 94% top-100 overlap. The algorithm is documented; the data inputs are free. This isn't a product yet — it's evidence that the firm's AI agents produce output at the level of paid incumbents.
Phase A validation
We held out two daily snapshots from the regression and computed the top-100 overlap between Empirica Score and a publicly-visible third-party scoring benchmark — both snapshots fall within the same two-week fit window. The pass threshold was set in advance at 60%, and both results comfortably cleared it. This measures how closely the open model reproduces an existing screen on free data — evidence of screening quality, not a forward-return or alpha claim (that is what Phase B tests).
Self-reported, pending external reproduction. These numbers come from our own validation run; no independent party has re-run the pipeline yet. The artifact-by-artifact package that will make that possible — code, data, protocol — is tracked openly at the reproducibility page.
Top-100 overlap
95.8%
2026-05-22
Top-100 overlap
94.9%
2026-05-12
Regression R²
0.859
800 obs · 8 snapshots
Monthly data cost
$0
vs $2,000+ for Bloomberg
RS proxy
passedPearson 0.94
Relative Strength against the universe. The strongest single proxy — anchors the score.
TA proxy
ceilingPearson 0.53
Technical signals. Structural ceiling without intraday data — moderately predictive.
FA proxy
ceilingPearson 0.61
Fundamental signals from public filings. A paid feed would lift this further.
What it would cost you another way
The Empirica Score is research evidence — not a product you can subscribe to today. The comparison below shows what it would cost to run an equivalent equity screen using the tools institutional buyers actually use.
| Option | Cost |
|---|---|
| Empirica Score | $0 / month data |
| Bloomberg Terminal | $2,000 / month |
| FactSet Workstation (institutional seat) | $1,000 – $2,000 / month per seat |
| Refinitiv Workspace | $1,800+ / month per seat |
| Retail screening service (Finviz Elite, similar) | $25 / month |
| Build it yourself | $50,000+ engineering + paid data feed |
Who this is for
Independent researcher
Roughly $24,000/year of seat cost avoided.
The likeliest case. You get a transparent, auditable score; a paid terminal gives you a closed box for several thousand percent more.
Wealth-management desk — daily top-100
$12,000 – $24,000/year saved per analyst seat.
Scales linearly with team size. A 5-analyst desk on FactSet costs $60,000 – $120,000/year just for the screener seats.
Quant team validating its own factor model
A defensible external benchmark you can actually cite.
The whole point of an open algorithm is replication. Closed scores can't be benchmarks for serious quant work.
Methodology
Three open factor proxies — Relative Strength (RS), Technical (TA), and Fundamental (FA) — were built on free inputs (yfinance price history, public sector classifications, public fundamentals). We ran an OLS regression of these proxies against a publicly-visible third-party scoring benchmark on 800 paired observations across 8 daily snapshots, finding R² = 0.859. The fitted weights are the Empirica Score. Two held-out snapshots gave 95.8% top-100 overlap on 2026-05-22 and 94.9% on 2026-05-12 — both well above the 60% pass target set in advance. Phase B (next) layers 6 weeks of shadow-trading and a 5-year out-of-sample backtest on top, after which the production weights are locked.
The formula at a glance
Empirica Score = 0.221·RS
+ 0.190·TA × 9.9
+ 0.307·FA × 9.9
+ 1.311 if NASDAQ-100
− 0.028 if S&P 500
− 0.492 if Russell 2000
− 6.914 if defensive sector
+ 0.220·log10(market cap)
+ 23.98
Sectors classified as defensive: Health Care,
Consumer Staples / Defensive, Utilities.
Weights fitted by OLS on 800 paired observations
across 8 daily snapshots (R² = 0.859, last refit
2026-05-26).The full source is in scripts/empirica_score/ in our repository. The three proxies (rs_proxy.py, ta_proxy.py, fa_proxy.py) are independently testable — and the regression that combines them is reproducible from any 8 daily snapshots.
Phase B · next
Phase A proves the algorithm reproduces a known benchmark. Phase B tests whether the same weights generalise across time and market regimes. The headline validation is a 5-year out-of-sample historical backtest using the same scoring rules applied to data the model has never seen. A separate live shadow-trading window on a hardened paper-trading setup adds forward-looking confirmation once we stand it back up.
Only when both pass do the weights lock for production use. Until then, the Empirica Score is research evidence — not an investment product. We share the methodology so peers can replicate it; we don't offer trading recommendations.
The current phase is stabilisation — a minimum of 90 clean days of operation on internal capital before any public trading-track-record dashboard ships, and any external-capital structure remains gated on a formal Australian financial-services lawyer review. We err toward over-disclosure of methodology and under-disclosure of live state — the reverse is the path most firms in this category get into trouble with.
The Empirica Score is one output of a fleet of always-on research agents. The rest — long-form publications, short notes, daily news synthesis, a sector-pulse macro signal — sits behind a subscription.