A signal-intelligence platform built to the standard your risk committee already enforces.

§ 01Why we wrote this document.

If you manage other people's money, your vendor stack lives or dies in a single meeting: the one where your CCO, CRO, and external auditor sit down with the head of the strategy desk and ask, line by line, "how do you know this number is right?"

Most signal-intelligence vendors fail that meeting. Their composite scores arrive without a credible interval. Their backtests don't disclose trial count. Their models change weekly with no version identifier. Their flow analytics can't be reproduced six months later because the inputs were never archived. When the inevitable drawdown forces a "what was the model saying that morning?" question, the vendor's answer is some variation of "we don't know."

DealerFlow Institutional was built for the other meeting. The one where you can produce, in two CLI commands, the model version, training-data hash, raw input snapshot, regime classification, calibrator unique ID, per-family attribution, and 95% credible interval that produced any historical prediction on file. Every prediction. Every day. Indexed by (symbol, timeframe, predicted_at). Reproducible from raw inputs forward, forever.

We assume your mandate documents already require independent validation, ongoing monitoring, model-change governance, and reproducibility. We built to that floor — not to a marketing brochure. — DealerFlow Institutional design contract

§ 02The three questions a portfolio manager actually asks.

"Can I show this to my IC tomorrow?"

Every prediction surfaces with a calibrated probability, a sample size N, a 95% credible interval, the prevailing regime classification, and a per-family attribution. We do not produce point estimates without uncertainty. When the cell is sparse, the surface shows INSUFFICIENT DATA rather than a fabricated number — the only honest default for a tool that goes in front of an investment committee.

"Can I attribute a bad month back to a vendor input?"

Yes. Open the audit trail panel, query the prediction window, and the Shapley-style decomposition tells you exactly which signal family contributed to which directional call, per day, per symbol. If dealer-flow positioning over-influenced a losing quarter, you can prove it. If technicals carried the alpha in a winning month, you can prove that too. The attribution is computed at prediction time and locked into an append-only table the moment the prediction is written.

"Can my CCO defend our vendor selection in a Form ADV brochure update?"

We supply, on request, a vendor-due-diligence pack that satisfies the substantive criteria most chief compliance officers apply for software analytics under Investment Advisers Act §206 and the Marketing Rule (Rule 206(4)-1): description of the service, methodology disclosure, conflict-of-interest statement, performance-claim framework, data-handling and confidentiality policies, business-continuity disclosure, insurance certificate, and SOC 2 readiness roadmap. The pack is delivered under NDA in week one of the pilot.

§ 03SR 11-7, mapped to capabilities your auditor can verify.

U.S. banking regulators codified the modern model-risk-management discipline in Federal Reserve SR 11-7 (April 4, 2011). The framework is not formally required of investment advisers, but sophisticated buy-side firms — pensions, endowments, multi-strategy hedge funds, and separately-managed-account allocators — increasingly apply its standards to any analytic model that influences a portfolio decision. We built to it because we believe the manager managing $100M for retirees should not lower the bar from the one their bank pension trustees already apply.

Each row below names a SR 11-7 requirement and the specific DealerFlow Institutional control your model-risk team can verify against logs and code:

The point is not that we cite the regulation. The point is that your model-risk team can read our code, our schema, and our logs, and verify each row above as an engineering claim rather than a marketing claim. Pilots include a guided walkthrough of the controls with your MRM lead in week one.

§ 04Methodology, with citations.

Every methodology choice in DealerFlow Institutional grounds in published academic or regulatory work. The list below is what your quant team should expect to vet during diligence; we supply working-paper references and our implementation files for any item:

• Purged k-fold cross-validation with embargo — López de Prado, Advances in Financial Machine Learning (Wiley 2018), Chapter 7. The only published CV methodology that handles serially-correlated labels without leakage.
• Deflated Sharpe Ratio — Bailey & López de Prado (2014), SSRN 2460551. Corrects headline Sharpe for selection bias, skewness, kurtosis, and sample size. Reported alongside raw Sharpe on every backtest.
• Probability of Backtest Overfitting — Bailey, Borwein, López de Prado, Zhu (2017), Journal of Computational Finance. Combinatorially-symmetric CV overfit detector. PBO reported on the Backtest Lab leaderboard.
• Hierarchical Bayesian partial pooling — Gelman & Hill (2007). Sparse cells borrow strength from related cells, producing better calibrated probabilities under low-N regimes. Implemented in PyMC.
• Isotonic probability recalibration — Niculescu-Mizil & Caruana (2005), ICML. Caveat acknowledged: needs N≥1000 per cell, otherwise Platt fallback. Applied in score-space after composite scoring.
• Shapley-value attribution — Lundberg & Lee (2017), NeurIPS, arXiv:1705.07874. Per-family contribution computed at prediction time, served by the diamond drill-down endpoint.
• Model Cards — Mitchell et al. (2019), ACM FAccT, arXiv:1810.03993. Auto-generated per champion. Live in your tenant.

This is the citation set a serious due-diligence team will recognize. It is not the only one in our codebase, but it is the one most relevant to the question "are these people using yesterday's methods?" They are not.

§ 05Capabilities a discretionary mandate manager actually uses.

The retail capability set is published on our public site. The capabilities below are the ones unique to the institutional tier — the ones that show up in your client reporting, your IPS compliance evidence, and your quarterly performance attribution:

Notably absent from the retail surface: the full audit-trail read API, model cards, reproducibility check daemon output, the SR 11-7 control register, and the regulator-CLI. Those are the controls your CRO will care about; they are gated to the institutional tier.

§ 06Operational posture for a discretionary fiduciary.

What we are.

A software-as-a-service analytics platform delivered over TLS. We display data, we compute signals, we maintain an audit log. We do not take discretion. We do not place orders. We do not custody assets. Our role ends at the screen.

What we are not.

Not a broker-dealer, not a registered investment adviser, not a futures commission merchant, not a custodian. We are not a fiduciary to your end clients; your firm is. We do not solicit trades or recommend specific securities to your principals. The contractual relationship is vendor-to-firm, not vendor-to-end-investor.

Data handling.

Tenant isolation at four layers: AWS network (private subnets), Cognito JWT, PostgreSQL row-level security (every query scoped to a verified tenant_id), and application-level RequestContext. Encryption: TLS 1.2+ in transit, AES-256 at rest (RDS, ElastiCache, S3). Audit log retention configurable to your firm's policy; default 7 years for prediction records, 2 years for raw input snapshots, lifetime for model registry rows.

Business continuity.

Hosted in AWS us-east-1. Multi-AZ failover available at the institutional tier. RDS automated snapshots nightly; point-in-time recovery within the retention window. Functional health monitor publishes a public status surface; institutional tenants receive direct alerting via configurable channels (PagerDuty, Slack, email).

Security & certifications roadmap.

Penetration test under engagement at pilot kickoff. SOC 2 Type II audit on the institutional code path planned for completion within twelve months of first institutional contract. Bug-bounty program available. Quarterly access review per institutional tenant. Sub-processor list, data-processing addendum, and security questionnaire provided in week one of the pilot.

§ 07Commercial terms.

Institutional tier engagements run on annual master service agreements. Per-seat pricing starts at $1,495 per seat per month (annual term, paid quarterly). Volume tiers begin at three seats. Multi-strategy desks managing multiple mandates from one parent fund can elect a desk-level licence at negotiated rates.

Every engagement begins with a 30-day diligence pilot. The pilot is fully featured: complete platform access, full audit-trail read, model-card delivery, regulator-CLI access, a working session with your MRM lead, your CCO, and your quant team. No trading-volume commitment, no integration commitment, no data-residency commitment during the pilot. If at the end of thirty days the platform has not earned its place on your stack, we shake hands and part on good terms.

Pilot conversion typically settles on annual term plus optional professional services for custom report delivery, sub-account structure, or workflow integration. Master services agreement, mutual NDA, and data-processing addendum are exchanged on first conversation.

§ 08How to begin diligence.

We do not take cold credit cards or self-service signups for the institutional tier; the fit is mutual. Every pilot starts with a 30-minute working call between our team and yours. Agenda for the first call is short:

1. Your mandate: AUM, number of accounts, broad strategy posture, current vendor stack.
2. Your governance posture: who validates models, who signs vendor selection, what reporting the IC expects.
3. The pilot scope: which workflows you'd test in thirty days, which controls your CRO wants to verify, what success looks like for both sides.

Mutual NDA executed before any non-public information changes hands. We bring our vendor-due-diligence pack to the second meeting.