Adaptive Tactical & Long-Horizon Allocation System
A regime-aware decision-support system for managing portfolio exposure under uncertainty — multitenant, governance-first, and designed so that abstention is a legitimate output, not a missing answer.
Explore ATLAS with sample portfolio data in a sandboxed demo environment. No account required. Demo sessions use delayed market data and include no live portfolio information.
ATLAS is a portfolio governance and macro-regime monitoring system designed to help investors manage exposure to uncertainty in global markets.
Rather than attempting to predict short-term price movements, ATLAS focuses on detecting structural regime shifts, systemic stress, and instability across financial markets. The system integrates multiple cross-asset signals to evaluate whether conditions favor risk-taking, caution, or abstention.
ATLAS is designed around a central principle: managing exposure to uncertainty is more important than maximizing returns during unstable regimes.
Instead, the system provides a structured framework for evaluating macroeconomic conditions and portfolio alignment across regimes.
ATLAS is built as a multitenant platform from the ground up. Every artifact, view, and decision surface is associated with an explicit tenant and operator context, so what one operator sees, exports, or acts on is scoped, attributable, and isolated from the rest of the system.
Classification of prevailing macro conditions from a cross-asset signal set, with confidence measurement, persistence analysis, and stability diagnostics. Designed to be slow-moving and interpretable.
Fail-closed detection of deteriorating classification certainty, producing persistence-aware decision states — Allocate, Reduce, and Abstain — with hysteresis controls to prevent flicker.
Abstention treated as a first-class decision output with a defined episode lifecycle, post-hoc outcome evaluation, trigger attribution, and regime-conditioned economic accounting.
Multi-horizon regime-transition diagnostics, including flip probability estimation, hazard-rate modeling, and persistence analysis. Published as a research-grade diagnostic signal; not coupled to exposure policy.
Continuous monitoring of decoupling between structural and tactical layers, with leakage detection, contamination audits, and tiered escalation. Ensures fast signals do not contaminate slow ones.
Multi-broker position ingest, exposure decomposition, concentration and regime-alignment analysis, target generation, and sleeve-scoped governance of proposed changes.
Version-controlled portfolio workflow with change history, merge governance, counterfactual replay under alternative regime paths, and auditable experimentation with hypothesis tracking.
Tail-protection workflow with premium budgeting, ladder construction, gap reconciliation, and position reconciliation. Protection cost is reported as an estimate proxy; real option-chain pricing is deferred.
Artifact freshness validation with canonical signature tracking, calibration and horizon-sensitivity audits, robustness stress testing, and full pipeline provenance logging.
Hypothesis registry with shadow-mode evaluation and promotion gating. Evidence-velocity tracking, scenario comparison, and validation requirements before any candidate change enters governance.
Layered explanation of each decision across structural, tactical, constraint, and availability dimensions, with gated disclosure of implications (allowed, qualified, or withheld).
Operator-scoped governance with isolated override state, per-operator sleeve management, composite-key entitlements, operator-aware audit trails, and fail-closed tenancy enforcement.
OIDC authentication, sleeve-based authorization with group mapping, entitlement system, full decision audit trail, override accounting, and governance logging.
Single-pass diagnostic baseline joining decisions, NAV-denominated realized economics, regime context, forward evidence, and posture flags. Provides an auditable end-to-end view of decision history and outcomes; runs as a measurement reference, not as allocation authority.
Decision-facing surfaces declare their evidence basis explicitly via decision-support and evidence-posture badges (live governance, shadow replay, diagnostic model, or insufficient). Per-layer registry enforces that no panel renders an implicit claim; missing inputs surface as labeled states, not silent fallbacks.
Most investment systems focus on maximizing returns under stable market conditions. ATLAS is designed for the opposite problem: navigating periods when markets become unstable, regimes shift, and traditional assumptions break down.
By emphasizing regime awareness, uncertainty management, and abstention as a legitimate decision state, ATLAS provides a framework for disciplined portfolio governance across changing macro environments.
ATLAS is in active production with continuous development. The production environment delivers governance-grade regime monitoring, portfolio management, and operator-isolated decision tooling. The beta environment provides access to experimental features and ongoing system improvements.
For pricing inquiries, institutional licensing, and operator onboarding, please visit the Contact page.
Release notes and development milestones
A production update is live at app.atlas-portal.ca. The release sharpens how ATLAS represents what it knows, what it does not know, and how the platform talks to operators about both.
A major production update is now live at app.atlas-portal.ca. The release brings ATLAS’s multitenant governance posture, pipeline reliability, and structural forecasting capabilities to production grade.
A new beta version is now available at beta.atlas-portal.ca incorporating significant new capabilities:
ATLAS Portal is live. The initial release introduces the core cross-asset regime classification model — a system for detecting structural regime shifts, systemic stress, and instability across global financial markets.
Formal model objects, governance, and admissible decision rules
ATLAS is specified as a tuple of measurable and governed objects. Every component is defined in the sections that follow with its domain, role, and admissible use.
ATLAS is a measurement and decision-governance system. It is not a return forecaster, an unconstrained optimizer, or an execution engine. Sections 2–14 define each component formally; Sections 15–16 give the public glossary and the explicit non-claims.
All random variables in this document are defined on a fixed probability space \((\Omega, \mathcal{H}, \mathbb{P})\) equipped with the discrete-time filtration \((\mathcal{F}_t)_{t \in \mathbb{Z}_{\ge 0}}\) introduced in Section 2. Statements of equality involving random variables are understood \(\mathbb{P}\)-almost surely; statements of equality involving deterministic objects are absolute.
The remainder of the document operates under the following labeled assumptions. Each is stated as part of the specification, not asserted as universally true of the underlying market or of the data-generating process.
Assumptions A1–A5 are referred to by label where their force is invoked.
Let \(t\) denote decision time and \(d\) an observation date. Let \(X_t \in \mathcal{X}\) denote the observable state vector and \(\mathcal{D}_t\) the set of artifacts available at \(t\):
For each artifact define its maximum observation date and lag:
Artifact state is a closed-set label:
Eligibility for downstream use admits labeled states (OK, STALE, DEGRADED) and blocks the two value-less states (UNAVAILABLE, INSUFFICIENT):
A STALE or DEGRADED artifact carries its label downstream rather than being silently substituted; an UNAVAILABLE or INSUFFICIENT artifact has no value to consume, and downstream computations fail closed (Section 10).
The natural artifact filtration is the increasing family of \(\sigma\)-algebras
where \(I_t\) is the set of artifact indices available at \(t\). By Assumption A2, every decision-time object defined in Sections 3–8 is \(\mathcal{F}_t\)-measurable: ATLAS cannot use information that is not in \(\mathcal{F}_t\), and any apparent use of such information is a governance violation.
Stale-state truthfulness. Define a current-day-complete artifact as one with \(\ell_i(t) = 0\) and \(q_i(t) = \mathrm{OK}\); a latest-available artifact as one with \(\ell_i(t) \ge 0\) and \(q_i(t) \in \{\mathrm{OK}, \mathrm{STALE}, \mathrm{DEGRADED}\}\). The two are not the same object; the latter must carry its label. Absence of data is not evidence of stability: \(q_i(t) = \mathrm{UNAVAILABLE}\) does not imply \(T_t = \text{Stable}\) (in fact, by A3, it forbids it). Artifacts are the measurement substrate from which all downstream objects in \(\mathcal{S}\), \(\mathcal{T}\), \(\mathcal{C}\), \(\mathcal{F}\), and \(\Pi\) are derived.
In the live implementation, \(q_i\) is decomposed into three orthogonal axes — a freshness label (timeliness and calendar-completeness), a health label (input quality), and an observability label (sample sufficiency for derivative quantities) — treated as governance vocabulary rather than analytic detail. The closed set above is the public abstraction.
The structural regime \(S_t \in \mathcal{S}\) is a finite-valued classification of the prevailing macro environment. Let \(Z_t = (Z_{1,t}, \ldots, Z_{k,t})\) denote a low-dimensional vector of governed scores along the growth, inflation, risk-appetite, and liquidity axes. Then
\(\Theta_S\) denotes the governed structural parameters; \(G_t\) is the artifact-admissibility state at time \(t\). The map \(f_S\) is stateful (the argument is the score history \(Z_{1:t}\), not \(Z_t\) alone); slowness and persistence of the resulting state sequence are formalized by the inequality below. Structural-regime estimation is in the regime-switching tradition (Hamilton, 1989), in which the latent state evolves as a Markov chain with transition matrix
The diagonal \(P_{ii}\) is the one-step self-persistence of regime \(i\); the off-diagonal mass governs reachability between regimes. \(P\) is a member of \(\Theta_S\) and is estimated on a declared calibration window (Section 14).
Structural regime is not tactical instability:
\(T_t\) does not appear as an argument of \(f_S\); equivalently, \(S_t\) is \(\sigma(Z_{1:t}, \Theta_S, G_t)\)-measurable and not \(\sigma(T_t)\)-measurable in general. A single-layer model that conflates the two horizons violates the persistence inequality above (Ang & Timmermann, 2012). Separation is monitored by the leakage diagnostic \(\mathrm{lk}(W) = I(S_t; T_t \mid Z_t)\) of Section 12 (HEDGEHOG).
The tactical state \(T_t \in \{\text{Stable},\, \text{Transitional},\, \text{Unstable}\}\) is derived from high-frequency diagnostics of classification certainty. Define discrete differences:
The tactical objects are:
The tactical fusion is
Thresholds and parameter values within \(\Theta_T\) are not disclosed. The fail-closed invariant is stated as an implication:
Equivalently: a tactical input that is NA-driven or whose governing artifacts are inadmissible can never resolve to Stable. The structural-tactical separation invariant — \(T_t\) does not appear as an argument of \(f_S\) (Section 3) — is enforced by construction. Any structural override is the subject of a separately governed rule, never implicit in tactical motion.
Confidence \(C_t \in \mathcal{C}\) is defined as proximity to a decision boundary (Bishop, 2006, Ch. 4), not as subjective belief. Let \(M_s : \mathbb{R}^k \to \mathbb{R}\) denote the score/margin function assigned to structural state \(s\). The live confidence object is the score-margin form:
Let \(\Gamma \subset \mathbb{R}^k\) denote the structural decision boundary — under the margin form, the indifference locus \(\Gamma = \{\,z \in \mathbb{R}^k \,:\, M_s(z) = M_{s'}(z) \text{ for some } s \neq s'\,\}\); under the geometric form, an estimated surface in the score space. A geometric distance formulation
is computed in a shadow geometry layer as a research-only diagnostic and does not gate live policy. In both formulations, the boundary limit
holds by construction. Both \(C_t\) and \(\Delta C_t\) are first-class \(\mathcal{F}_t\)-measurable quantities (A2); \(\Delta C_t < 0\) is information about the trajectory and is not equivalent to a regime flip.
Flip risk \(F_t \in \mathcal{F}\) is a function of proximity and motion in the score space:
The map \(\phi\) is monotone-non-increasing in \(C_t\) (proximity to the boundary raises flip risk) and increasing in \(\|\Delta Z_t\|\) (velocity in the score space raises flip risk). Three properties hold by construction:
Score disagreement is the range across the structural-score axes (the live implementation applies the max–min functional on a designated axis subset):
\(D_t^{\,score}\) is a live object and is one of the families that contributes to admissibility of an abstention (Section 8).
Surface disagreement is the indicator that eligible classification surfaces disagree on the structural label:
\(D_t^{\,surface}\) is research-only; it is measured but not wired into the live admissible-action map. Any promotion to live status would require an explicit, declarative governance pass (Section 10).
The policy map is
where \(\mathcal{U}\) is the set of admissible non-null exposure actions and \(\varnothing\) denotes no action — abstention. In the current implementation:
The admissible decision set at time \(t\) is the gated subset
where \(u\) is the admissibility map and \(\pi_t\) is \(\mathcal{F}_t\)-measurable by A2. Gating reduces the action set; it does not select within it.
Abstention is not failure. It is a first-class output with reason codes, an episode lifecycle, and governance attribution. ATLAS manages exposure to uncertainty; it does not maximize return. This formulation follows the reject-option framework of Chow (1970) and its extension to selective classification in El-Yaniv & Wiener (2010).
An abstention episode is a maximal time interval over which \(\pi_t = \varnothing\), paired with a post-hoc outcome label. Formally:
where the outcome label \(\omega\) is assigned post-hoc at a pre-registered forward horizon:
Horizons and drawdown thresholds are declared in advance to prevent horizon-selection bias (Harvey, Liu & Zhu, 2016). \(\mathcal{E}^{\,\mathrm{abst}}\) is the input set to the LANTERN confusion matrix (Section 12).
ATLAS enforces four disjoint layers. The boundaries are architectural, not stylistic; each is stated as a measurability or domain identity.
Let \(\mathcal{V} = \{V_1, V_2, \ldots, V_n\}\) be the family of validators, each producing a closed-set verdict:
Fail-closed semantics:
The validator family covers, at minimum:
A \(\mathrm{FAIL}\) verdict is non-recoverable: governance refuses to publish. \(\mathrm{INSUFFICIENT}\) expresses a well-defined computation against a sample too thin to commit a value — it is labeled, not silently coerced to Stable or to a default. Governed roots are never written from shadow tracks. Hypotheses that fail empirical validation are recorded and disabled (Popper, 1959).
Let \(o \in \mathcal{O}\) denote an operator and \(s \in \mathcal{S}_o\) a sleeve in that operator’s scope. Authorization is a closed-set predicate over \((o, s, r)\) where \(r\) is the requested resource:
A request is valid only if
Missing operator context is fail-closed: a surface that requires \((o, s)\) to be safe refuses to render rather than render against a default. Let \(\mathrm{scope}(D)\) return the \((o, s)\) tuple to which an artifact is scoped, with \((o, \varnothing)\) marking org-wide artifacts. The per-operator-sleeve filtration is
Cross-sleeve leakage is structurally prohibited: any value rendered for \((o, s)\) is \(\mathcal{F}_t^{(o,s)}\)-measurable, and \(\mathcal{F}_t^{(o,s)} \cap \mathcal{F}_t^{(o,s')}\) for \(s \neq s'\) contains only org-wide \((o, \varnothing)\) state. Reports, exports, overrides, governed artifacts, and promotion gates are all scoped to \((o, s)\). The audit trail records \(\mathrm{scope}(\cdot)\) for any action taken, so post-hoc review can reconstruct what was observable when a decision was made.
The programs below are measurement and evaluation tracks, each with an explicit status tag. None of them are allocation engines on their own; promotion to live policy requires the gates of Section 10.
HEDGEHOG STATUS: LIVE. A forensic-measurement program for the regime-and-tactical stack. Its scope is the tuple of measurement axes
Each axis admits a formal definition over a discrete window \(W = [t_0, t_1] \cap \mathbb{Z}_{\ge 0}\) and a scope \(R\) (a subset of artifacts, regimes, or sleeves):
Reading the axes:
HEDGEHOG is a measurement program, not an allocation surface. None of \(\mathrm{cov}\), \(\mathrm{lat}\), \(\mathrm{per}\), \(\mathrm{cal}\), \(\mathrm{lk}\) enters \(\mathcal{U}_t\). Instrumentation depth along each axis is itself a HEDGEHOG-reported metric, not a uniform claim of completeness.
SENTINEL STATUS: RESEARCH. Transition-forecasting evaluation. Let \(\hat{p}^{(M)}_{ij}(t; h)\) denote model \(M\)'s estimate of \(\mathbb{P}(S_{t+h} = j \mid S_t = i)\). For a horizon \(h\), a scoring rule \(\ell\), and an out-of-sample index set \(\mathcal{I}_h\):
SENTINEL reports \(\Delta\mathcal{L}(h)\) over a held-out window and across horizons \(h\), comparing a baseline \(B\) to a candidate \(C\). There is no map from \(\Delta\mathcal{L}\) into \(\mathcal{U}_t\): SENTINEL is measurement of forecastability, not policy.
LANTERN STATUS: SHADOW. Caution-policy calibration. Let \(\pi_t^{\star}\) denote a post-hoc reference action evaluated against forward outcome at a pre-registered horizon (see Section 8). LANTERN computes the policy-state confusion matrix over the realized policy \(\pi_t\) and the reference \(\pi_t^{\star}\):
\(\Lambda\) is reviewed by operators in a shadow learning surface. LANTERN does not write into \(\Pi\); there is no \(\Lambda\)-derived map into \(\mathcal{U}_t\).
TICM STATUS: LIVE labeling. Stress-event window labeling. TICM defines a family of event windows and a corresponding indicator:
\(\tau_t^{\,\mathrm{TICM}}\) participates as a conditioning variable in the evaluation programs above. TICM labels never enter \(\mathcal{U}_t\) directly and carry no policy authority of their own.
WS7 STATUS: RESEARCH. Surface-disagreement track. Building on \(D_t^{\,surface}\) (Section 7), WS7 reports the windowed disagreement rate
and stratifications of it by structural regime \(S_t\) and by TICM label \(\tau_t^{\,\mathrm{TICM}}\). WS7 is measured but not wired into \(\mathcal{U}_t\); there is no \(\rho^{\,\mathrm{WS7}}\)-derived gate on \(\pi_t\).
A protection bundle at decision time \(t\) is a finite set of contracts
with strike grid \(\{k_j\}\), contract counts \(\{n_j\}\), and per-contract proxy premia \(\{c_j\}\). The protection-cost proxy is
where \(c_j^{(t_0)}\) is the contract-proxy premium taken at a snapshot date \(t_0\) (\texttt{STATIC\_PROXY\_AT\_T0}) — not the live mid-market price at \(t\). The budget constraint is
The Slice F successor would replace each \(c_j^{(t_0)}\) by a live mid-market \(c_j(t)\) drawn from an option-chain pricing surface; this remains DEFERRED. Until promotion, \(P_t^{\,\mathrm{proxy}}\) carries no allocation, abstention, or gross-reduction authority and does not equal the realizable cost of acquiring \(\mathcal{P}_t\) at \(t\). It is a reporting estimate, labeled as such at the row level.
Only methods that are implemented or explicitly evaluated against ATLAS data are described here. ATLAS does not claim methods it does not use.
Structural-regime estimation is in the regime-switching tradition (Hamilton, 1989) with a finite latent state and a transition matrix \(P\) (Section 3). Parameters \(\Theta_S\) and \(\Theta_T\) are estimated on a calibration window declared in advance and held out from out-of-sample evaluation:
For a generic parameter \(\theta \in \Theta\):
Sample-size guardrail. Estimators carry an explicit minimum-sample requirement \(n_{\min}(\theta)\); below threshold, the estimator returns a labeled non-value rather than committing one:
The following properties are encoded in the formal objects above and in Sections 2–8:
Parametric confidence intervals, bootstrap-based interval estimation, and spline-based smoothing are not claimed as components of the live estimation stack. Parameter uncertainty is real and is not collapsed into false precision.
Every cross-section ATLAS symbol introduced in Sections 1–14 and 16 appears below with its domain, interpretation, implementation status, and a public-safety note. Standing assumptions A1–A5 are stated in Section 1.1. Standard set-theoretic and probability notation is not duplicated. Program-local symbols defined inline within a single subsection — the abstention-episode set \(\mathcal{E}^{\,\mathrm{abst}}\) and outcome labels \(\{\mathrm{JC}, \mathrm{FC}, \mathrm{IN}\}\) (Section 8); the scope predicate \(\mathrm{scope}(\cdot)\) and per-operator-sleeve filtration \(\mathcal{F}_t^{(o,s)}\) (Section 11); the structural decision boundary \(\Gamma\) (Section 5); the HEDGEHOG axes \(\mathrm{cov},\, \mathrm{lat},\, \mathrm{per},\, \mathrm{cal},\, \mathrm{lk}\) and supporting \(W,\, R,\, B_k,\, \tau_s\); the SENTINEL objects \(\hat{p}^{(M)},\, \mathcal{L}^{(M)},\, \Delta\mathcal{L},\, \mathcal{I}_h\); the LANTERN objects \(\pi_t^{\star},\, \Lambda,\, M_{\pi,\,\pi^{\star}}\); the TICM objects \(\mathcal{E}^{\,\mathrm{TICM}},\, W_e,\, \tau_t^{\,\mathrm{TICM}}\); the WS7 statistic \(\rho^{\,\mathrm{WS7}}\); the optionality objects \(\mathcal{P}_t,\, k_j,\, n_j,\, c_j,\, B_t\); the calibration objects \(W_{\mathrm{cal}},\, W_{\mathrm{oos}},\, n_{\min}\); and the admissibility map \(u\) — are not duplicated here. Implementation status is one of LIVE, SHADOW, RESEARCH, or DEFERRED.
| Symbol | Domain | Interpretation | Status | Public-safe note |
|---|---|---|---|---|
| \(t\) | \(\mathbb{Z}_{\ge 0}\) | decision time index | LIVE | — |
| \(d\) | calendar dates | observation date | LIVE | — |
| \(X_t\) | \(\mathcal{X} \subseteq \mathbb{R}^p\) | observable state vector | LIVE | — |
| \(D_i\) | \(\mathcal{D}\) | governed artifact \(i\) | LIVE | schema / keys / dates validated |
| \(G_t\) | \(\mathcal{G}\) | artifact-admissibility state | LIVE | fail-closed |
| \(\mathcal{F}_t\) | \(\sigma\)-algebra | natural artifact filtration at \(t\) | LIVE | distinct from flip risk \(F_t\) (calligraphic vs italic) |
| \(Z_t\) | \(\mathbb{R}^k\) | structural-score vector | LIVE | growth / inflation / risk / liquidity axes |
| \(S_t\) | finite \(\mathcal{S}\) | structural regime | LIVE | slow, persistent |
| \(T_t\) | \(\{\text{Stable},\, \text{Transitional},\, \text{Unstable}\}\) | tactical instability state | LIVE | fast, fail-closed |
| \(P\) | \([0,\,1]^{|\mathcal{S}|\times|\mathcal{S}|}\), row-stochastic | structural transition matrix | LIVE | diagonal = self-persistence; member of \(\Theta_S\) |
| \(\Theta_S\) | parameter space of \(f_S\) | governed structural parameters | LIVE | values undisclosed; calibrated on declared window |
| \(\Theta_T\) | parameter space of \(f_T\) | governed tactical parameters | LIVE | values undisclosed |
| \(H_t\) | \(\mathbb{Z}_{\ge 0}\) | transitional severity | LIVE | tactical scalar; thresholds undisclosed |
| \(N_t\) | \(\{0,\, 1\}\) | NA-driven indicator | LIVE | routes to Transitional when set |
| \(K_t\) | \(\mathbb{Z}_{\ge 0}\) | trigger count | LIVE | number of active triggers at \(t\) |
| \(C_t\) | \(\mathbb{R}_{\ge 0}\) | confidence (margin form) | LIVE | geometric form is SHADOW |
| \(F_t\) | \(\mathbb{R}\) | flip-risk / transition pressure | LIVE | acceleration-aware |
| \(\Delta\) | operator | first difference | LIVE | — |
| \(\Delta^2\) | operator | second difference | LIVE | — |
| \(D_t^{\,score}\) | \(\mathbb{R}_{\ge 0}\) | score-axis dispersion | LIVE | contributes to abstention admissibility |
| \(D_t^{\,surface}\) | \(\{0,\, 1\}\) | surface-label disagreement | RESEARCH | not policy-wired (WS7) |
| \(\pi_t\) | \(\mathcal{U} \cup \{\varnothing\}\) | policy / exposure posture | LIVE | no execution authority |
| \(\varnothing\) | distinguished element of \(\Pi\) | abstain (no action) | LIVE | first-class output |
| \(V_j\) | \(\mathcal{D} \to \{\mathrm{PASS},\, \mathrm{FAIL},\, \mathrm{INSUFFICIENT}\}\) | validator | LIVE | fail-closed on FAIL |
| \(E(D_i)\) | \(\{0,\, 1\}\) | artifact eligibility | LIVE | composite of schema / keys / dates / state |
| \(o\) | \(\mathcal{O}\) | operator | LIVE | missing \(\Rightarrow\) fail-closed |
| \(s\) | \(\mathcal{S}_o\) | sleeve in operator scope | LIVE | no cross-sleeve leakage |
| \(\operatorname{Auth}\) | \((o, s, r) \to \{0,\, 1\}\) | authorization predicate | LIVE | closed-set; logged |
Programs (Section 12) carry their own tags: HEDGEHOG LIVE, TICM LIVE labeling, LANTERN SHADOW, SENTINEL RESEARCH, WS7 RESEARCH. The optionality proxy \(P_t^{\,\mathrm{proxy}}\) (Section 13) is ESTIMATE_PROXY; real option-chain pricing is DEFERRED.
Three statements, made explicitly to prevent inferential drift.
ATLAS does not assert
as a universal trading claim.
ATLAS does not define
as an unconstrained return-maximization problem.
ATLAS does define
The objective ATLAS is governing is exposure to uncertainty under regime instability, subject to validation \(\mathcal{V}\), operator scope \((o, s)\), and the layer separation of Section 9. It is not portfolio-return maximization under regime stability. The two problems are complementary, not interchangeable.
Ang, A. & Timmermann, A. (2012). Regime changes and financial markets. Annual Review of Financial Economics, 4, 313–337.
Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer.
Chow, C. K. (1970). On optimum recognition error and reject tradeoff. IEEE Transactions on Information Theory, 16(1), 41–46.
Diebold, F. X. & Rudebusch, G. D. (1996). Measuring business cycles: A modern perspective. Review of Economics and Statistics, 78(1), 67–77.
El-Yaniv, R. & Wiener, Y. (2010). On the foundations of noise-free selective classification. Journal of Machine Learning Research, 11, 1605–1641.
Hamilton, J. D. (1989). A new approach to the economic analysis of nonstationary time series and the business cycle. Econometrica, 57(2), 357–384.
Harvey, C. R., Liu, Y. & Zhu, H. (2016). … and the cross-section of expected returns. Review of Financial Studies, 29(1), 5–68.
Popper, K. R. (1959). The Logic of Scientific Discovery. Hutchinson.
Taleb, N. N. (2007). The Black Swan: The Impact of the Highly Improbable. Random House.
Tetlock, P. E. & Gardner, D. (2015). Superforecasting: The Art and Science of Prediction. Crown.
Ask about ATLAS