Infrastructure / Signal Intelligence

The Signal Bank: 140+ Measured Edges

Every signal we've ever discovered, measured, and monitored — across equities, futures, crypto, and currencies. Stored in a 93-million-row data lake that grows every night.

8 min read Infrastructure Deep Dive June 2026
140+
Unique signals
93M
Rows in DuckDB
4
Asset classes covered
24h
Nightly automated updates
Section A

What is a signal bank?

A signal is any measurable quantity that has predictive power over future returns. Some signals are obvious — a stock that has risen for six months tends to keep rising. Some are subtle — the gap between what a CEO says and how they say it predicts performance six months later. Some are structural — the cost of holding a futures contract tells you something about supply and demand that the spot price alone cannot.

The signal bank is every signal we have ever discovered, measured, and tracked. Not a list of ideas — a live database. Every signal has been calculated across every instrument in our universe, every day, for as far back as the data exists. The result is 93 million rows of measured signal history stored in a DuckDB data lake that updates automatically every night.

Most signals are weak individually. An information coefficient of 0.03 to 0.08 is typical — barely above noise. The power comes from combining uncorrelated signals. The signal bank is the raw material. The portfolio engines are the finished product.

Section B

Eight categories, 140+ signals

Every signal falls into one of eight categories. Each category asks a different question about the market. The categories are not decorative — they reflect genuinely different sources of information, which is why signals across categories tend to be uncorrelated.

Momentum
~20 signals
"Does the trend persist?"
Price momentum at multiple horizons, from 21-day to 252-day. Cross-sectional momentum, time-series momentum, and momentum acceleration. The oldest and most documented anomaly in finance — assets that have been rising tend to keep rising.
mom_21d mom_63d mom_126d mom_252d mom_acceleration ts_momentum xs_momentum
12 deployed across Sentinel, Harvest, Meridian engines · 8 measured, not deployed
Spring
~15 signals
"Has the move gone too far?"
Z-scores relative to moving averages, distance from fair value, Hurst exponent, and reversion speed. When an asset deviates too far from its equilibrium, the probability of reversal increases — these signals measure how far is "too far."
zscore_20d zscore_60d dist_from_ma hurst_exponent reversion_speed bollinger_width
10 deployed in Spring engine · 5 measured, monitoring decay
Volatility
~15 signals
"Is the market calm or stressed?"
Average true range, realized vs implied volatility, vol-of-vol, vol ratio across timeframes, and regime classification. Volatility is not risk — it is information. These signals tell the system when to press and when to protect.
atr_14d vol_ratio realized_vs_implied vol_of_vol vix_term_slope garch_forecast
11 deployed in regime detection + Sentinel engine · 4 measured, not deployed
Carry
~10 signals
"What's the cost of holding?"
Funding rates, futures term structure, dividend yield, roll yield, and interest rate differentials. Carry is the oldest premium in finance — the return you earn simply for bearing a position. These signals identify when carry is rich enough to harvest.
funding_rate term_structure roll_yield div_yield carry_momentum
6 deployed in crypto carry + commodity engines · 4 measured, parked
Sentiment / NLP
~10 signals
"What are humans saying between the lines?"
CEO psychology scores from earnings transcripts, earnings surprise magnitude, analyst estimate dispersion, and management language patterns. The strongest single signal category in our arsenal — IC 0.115, t-stat 7.08 on the composite.
creator_energy adaptation_speed conviction_depth nlp_composite earnings_surprise estimate_dispersion
7 deployed as equity filter + standalone · 3 measured, expanding coverage
Cross-Asset
~15 signals
"How are markets connected right now?"
Equity-bond correlation, crypto-equity correlation, VIX term structure, credit spreads, and cross-market momentum. Markets are a network. When correlations shift, the regime is shifting — these signals detect the structural changes before they show up in returns.
eq_bond_corr crypto_eq_corr vix_term_structure credit_spread cross_momentum regime_score
9 deployed in regime allocator · 6 measured, used for monitoring
Microstructure
~10 signals
"What's happening under the surface?"
Order flow imbalance, volume profiles, overnight gaps, bid-ask spread dynamics, and dark pool activity. The footprint of institutional money moving before the headline. Most of these require intraday data — some are deployed, many are measured and waiting for higher-frequency infrastructure.
order_flow volume_profile overnight_gap spread_dynamics dark_pool_pct
3 deployed in execution layer · 7 measured, awaiting intraday pipeline
Fundamental
~15 signals
"What does the balance sheet say?"
Price-to-earnings, book-to-market, Piotroski F-score, analyst estimate revisions, and earnings quality. Classic value signals that have driven returns for decades. Individually weak on daily timescales, but powerful as regime-conditional filters over quarters.
pe_ratio book_to_market f_score analyst_revisions earnings_quality accruals_ratio
5 deployed as screening filters · 10 measured, used in research pipeline
Section C

Signal lifecycle

A signal goes through five stages before it can touch capital. Most signals die in stage two. The ones that survive the full pipeline are battle-tested — measured, combined, deployed, and continuously monitored for decay.

1
Discovery
A hypothesis emerges from research, the autonomous factory pipeline, or academic literature. At this stage it is nothing more than an idea — "does overnight gap size predict next-day returns?" The factory generates hundreds of these automatically every week.
2
Measurement
Calculate the information coefficient (IC), t-statistic, and decay rate across the full history. A signal needs a t-stat above 3.0 to be considered significant. Most hypotheses die here — they looked promising in a narrow window but fail on the full dataset.
3
Composition
Related signals are combined via walk-forward logistic regression. The composite is always stronger than any individual signal because it captures different facets of the same phenomenon. The NLP composite (3 signals into 1) is the clearest example — IC jumped from 0.06 individually to 0.115 combined.
4
Deployment
Wire the composite signal into the portfolio engine. The signal is assigned to one or more strategy legs, position sizing is calibrated, and the signal begins influencing live allocation decisions. Only signals that pass all three prior stages reach this point.
5
Monitoring
Track every deployed signal for decay, correlation drift, and regime sensitivity. If a signal's rolling IC drops below its historical floor for two consecutive months, the system automatically reduces its weight before capital is at risk. Signals are not set-and-forget — they are living instruments.
Section D

Why weak signals create strong portfolios

The combination principle

Most signals are weak individually. An IC of 0.03 to 0.08 is typical — barely above noise when measured on a single instrument over a single time horizon. If you showed any one signal to a traditional portfolio manager, they would dismiss it as statistically insignificant.

But portfolio mathematics does not care about individual signal strength. It cares about the correlation between signals. Ten weak signals that are uncorrelated with each other produce a composite that is dramatically stronger than any one of them alone. The signal bank is not a collection of edges — it is the raw material for constructing edges that no single signal could produce.

"The signal bank is the raw material. The engines are the finished product. No individual signal is the edge — the combination is."

Decay monitoring

Every deployed signal is monitored continuously for decay. Quantitative edges erode over time as more capital chases them. The system tracks each signal's rolling information coefficient against its historical baseline and flags any signal that drops below two standard deviations of its long-run average.

When a signal is flagged, its weight is reduced automatically — before it costs real money. The system does not wait for a drawdown to tell it something stopped working. It measures the signal's predictive power directly and adjusts in real time. This is the difference between a static model and a living system.

The signal bank grows every night. New data arrives, new signals are calculated, and the 93-million-row lake gets a little deeper. Some of these signals will eventually be promoted from "measured" to "deployed." Most will remain in the bank as insurance — measured, monitored, and ready to be activated if the market regime shifts to one where they become valuable.

The bank is not the strategy. It is the inventory that makes the strategy possible. And unlike physical inventory, measured data never expires. Every row we add today makes the system smarter tomorrow.

← Previous
The Research Funnel
Next →
The Factory Pipeline