That London fund and its rival made the same mistake from opposite directions. One side believes LLMs and foundation models will eventually replace factor models, stochastic processes, optimizers, and symbolic finance infrastructure. The other still treats AI as cosmetic — useful for memos, maybe transcript summaries, but irrelevant to real portfolio construction. Both views are obsolete.
The firms compounding edge in 2026 are not choosing between AI and classical quant. They assign tasks across a hybrid stack: LLMs handle semantic extraction, time-series foundation models handle temporal forecasting, classical optimizers make final decisions, symbolic compute guarantees numerical correctness, and humans govern the system and retain accountability. This is not a diplomatic middle ground — it is a hard architectural position. Every layer exists because a different class of models wins on a different failure surface.
Why this debate still matters.
Most firms do not fail because they picked bad models. They fail because they embedded category errors into the stack — asking language models to perform deterministic numerical functions, or asking static factor models to extract signal from a document-heavy market structure that has become increasingly semantic, fragmented, and latency-sensitive.
This matters because production quant is not an academic benchmark. In production, a false positive is not just a bad prediction — it is a position. A hidden temporal leak is not a modelling artefact — it is a live capital allocation error. An opaque reasoning chain is not merely inelegant — it becomes an audit failure the moment compliance asks why the system accumulated concentrated exposure into an event window. The reason hybrid architecture is winning is simple: it reflects the actual topology of the problem.
Markets are partly numerical, partly semantic, partly temporal, partly operational, and always adversarial. No single model class spans all of those dimensions well enough to own the stack.
Where pure AI breaks.
Five pure-AI failure modes recur in production:
- Numerical hallucination. Claude Opus 4.7, GPT-5.5 Pro, and DeepSeek V4 score above 95% on academic math benchmarks yet show 16–28% hallucination rates on realistic financial numerical tasks. A VaR engine off by 23%, a margin requirement rounded wrong, contract multipliers silently mixed — mistakes made in the only place a mistake is unacceptable: arithmetic tied to risk and capital.
- Temporal semantics collapse. Frontier LLMs leak future information into backtests 35–52% of the time when asked to reason about as-of states. As-reported vs. as-restated financials get conflated; revisions are read with hindsight. Backtest Sharpe of 2.1 becomes live Sharpe of 0.4, misdiagnosed as a regime shift when look-ahead bias was engineered into the research loop.
- Low-SNR signal extraction. True alpha lives where useful signal is 10–30% of variance; LLMs were pre-trained on high-SNR corpora. A sentiment engine emits "strong buy" on 34% of calls and the information coefficient lands at 0.02 — the noise floor. Fluency is mistaken for edge.
- Regime blindness. LLMs lack an intrinsic representation of volatility, correlation, liquidity, or microstructure state. A model tuned on 2019–2023 keeps emitting mean-reversion calls through a Q1 2026 volatility spike. The failure is quiet: Sharpe contribution turns negative for six consecutive weeks before intervention.
- The explainability wall. Reasoning traces of 15K, 60K, or 200K tokens are generated narratives, not explanations. FINRA, FCA, and MiFID II do not accept "the model considered many factors" as an audit trail. Compliance flags the strategy, risk asks why the concentrated position was built ahead of an event, the firm cannot answer — strategy suspended.
Where pure classical breaks.
If pure-AI fails by overextending probabilistic language systems into deterministic domains, pure-classical fails by underestimating how much of today's alpha surface is semantic, unstructured, and operationally too broad for humans to process fast enough. Five equally serious modes:
- Semantic blindness. A factor model cannot see the sentence on page 31 that changes guidance quality, nor detect management language shifting from confidence to evasive caution while headline numbers stay optically stable. If a competitor extracts 80 bps/year from tone revision and filing deltas, that edge is left on the table.
- Scalability wall on alternative data. A 3–5 person team processes ~30 documents/day with discipline; a modern semantic layer processes 400+. Classical firms sit on $2.4M/year data estates while processing only ~8% of the available signal. The bottleneck is scalable machine reading, not data acquisition.
- Strategy development velocity. Classical pipelines run 6–14 weeks hypothesis-to-deploy; agent-augmented competitors compress research and iteration into 3–5 days. Validate a card-spend signal a month after three faster rivals traded it, and the edge evaporated before the validation memo was finished.
- Static model decay. Factor models retraining semi-annually fall behind non-stationary markets. A value framework calibrated on 2021–2023 misses how the 2026 rate environment altered the growth-duration trade-off. The model is operationally too slow to track the state variables that dominate realised returns.
- Human bottleneck on monitoring. A 50,000+ position book with derivatives overlays and intraday crowding cannot be safely monitored by a person glancing at dashboards every 15–30 minutes. Liquidity holes and concentration spikes accumulate in under 20 minutes. Human judgment should govern the stack — not be its primary real-time monitor.
The five-layer hybrid stack.
The design principle is brutally simple: allocate each task to the component class that wins on that task under production constraints — latency, auditability, temporal integrity, transaction costs, error tolerance. Not vendor demos, not benchmark theatre.
LLMs — where they actually win. 88–94% on filing deltas.
Transcript sentiment, SEC filing delta detection, guidance-quality scoring, regulatory language shift. High-SNR semantic tasks close to what LLMs were trained for. 85–92% on transcript sentiment, 88–94% on 10-K change detection — via tool-augmented extraction with strict JSON schemas, confidence intervals, and weekly human review on a randomised <5% sample. Semantic parsers, not financial calculators.
Specialised sequence models. Distributions, not point forecasts.
Return forecasting across 1/5/21-day horizons, volatility regime detection, correlation shift, liquidity stress. TimesFM v2, Chronos-T5, and finance-specific foundation models preserve ordering, support causal masking, and emit probabilistic outputs. They feed the optimizer as uncertainty-weighted alpha signals — not deterministic commands.
Deterministic math still rules. Agents advise; optimizers decide.
Mean-variance, Black-Litterman, risk parity, factor neutrality, transaction-cost modelling, exposure constraints. Structured problems reward solvers that exploit convexity, duality, KKT conditions. The optimizer is deterministic, auditable, explainable — and never receives raw LLM output; semantics convert into structured alpha vectors with calibrated confidence and provenance first.
Non-negotiable for trade-affecting math. LLMs never do the arithmetic.
Every numerical operation tied to risk, sizing, execution, accounting, or valuation routes through typed compute — VaR, CVaR, Greeks, margin, P&L attribution, corporate-action normalisation. The LLM may generate a computation plan; NumPy, pandas, SymPy, and QuantLib execute it. Result: zero arithmetic hallucination on trade-affecting calculations.
Above all models. Strategy itself is a judgment problem.
Humans define what inefficiency is worth pursuing, decide which metrics matter, distinguish genuine regime change from noise, and own legal and reputational acceptability. Override, circuit breakers, model rollback authority, governance checkpoints — most important precisely when novelty enters. No institution can delegate accountability for live capital to an agent.
LLMs never do arithmetic that affects a trade, a risk metric, a margin figure, or an execution decision. A semantic layer can tolerate classification error if the optimizer shrinks exposure and oversight catches anomalies. A symbolic compute layer cannot tolerate hallucination, because its outputs directly transform capital, leverage, and control limits. Agents advise. Optimizers decide. Symbolic engines compute. Humans govern.
The architecture, data to decision.
Draw the stack as a pipeline. L1 ingests earnings calls, filings, analyst notes, card data, satellite and supply-chain text, emitting sentiment scores, guidance-quality tags, revision flags and confidence intervals. L2 ingests prices, returns, realised vol, options flow and cross-asset state, emitting expected-return distributions, regime probabilities and liquidity-stress estimates.
Both feed an alpha fusion layer: semantic outputs become scored features, time-series forecasts become uncertainty-aware signals, normalised and shrinkage-adjusted into a common format. The L3 optimizer maximises expected utility subject to exposure budgets, transaction costs and risk limits, producing target weights. A deterministic execution layer (VWAP, TWAP, POV, implementation-shortfall) routes the trades. L4 symbolic compute sits across all of it — VaR is not estimated by language, margin is not inferred by prose, Greeks are not approximated by verbal reasoning. L5 governance sits above the whole system: humans approve model updates, inspect anomalies, and pull circuit breakers when conditions move outside approved envelopes. It is not "AI plus quant." It is a layered control system.
How the hybrid stack compares.
| Task | Pure LLM | Pure classical | Hybrid |
|---|---|---|---|
| Earnings transcript processing | 87% accuracy | 0% (cannot read text) | 88% |
| Return forecasting · 5-day | 14–18% hit rate | IC ~0.04 | IC ~0.08 |
| Portfolio optimization | unusable | industry standard | industry standard |
| VaR calculation | 16–28% error | exact | exact |
| Alt-data processing velocity | 400 docs/day | 30 docs/day (human) | 400 docs/day |
| Regulatory explainability | poor | full | full |
| Strategy development cycle | 3–5 days | 6–14 weeks | 3–5 days |
| Regime shift detection | structurally weak | manual, days | <1 hour |
The point is not that hybrid wins by using "more technology." It wins because each layer is specialised. Hybrid stacks beat pure systems the same way good desk architecture beats ad hoc tooling — by decomposing the problem correctly.
What Nyquist built.
Nyquist was built around this logic from day one — a five-layer hybrid architecture in production, not a monolithic AI wrapper over legacy market data tooling. Semantic processing is handled by 36 specialised agents deployed through NVIDIA NIM microservices on L40S and A100 infrastructure: DeepSeek V4-Flash where throughput dominates, Claude Opus 4.7 where orchestration and complex semantic decomposition matter. Zero agents perform arithmetic directly.
Every numerical operation routes through a symbolic compute layer built on NumPy, pandas, QuantLib, and internal typed interfaces. Time-series forecasting is separated from semantic inference and handled by dedicated temporal models, including a proprietary 13B financial foundation model training for Q3 2026 on ~15TB of tick data and SEC filings. The portfolio decision is made by a classical optimisation layer using cvxpy and a proprietary risk engine.
The architecture is explicit: agents advise, optimizers decide, symbolic engines compute, humans govern. The operating profile is what matters — ~87% accuracy on semantic tasks, exact numerical precision on risk-critical calculations, IC ~0.08 on return forecasts from the specialised time-series layer, and p95 end-to-end latency <450ms across the full stack.
The assumptions behind these numbers, stated plainly:
- Model universe (May 2026): Claude Opus 4.7, GPT-5.5 Pro, Gemini 3.1 Pro, DeepSeek V4-Pro/Flash, Llama 4, Qwen3.6.
- Classical methods: Markowitz mean-variance, Black-Litterman, Almgren-Chriss execution, Grinold-Kahn for mapping signal to expected performance.
- Time-series models: TimesFM v2, Chronos-T5, Nyquist's proprietary 13B model under training.
- Evaluation criteria: information coefficient, Sharpe stability, regime robustness. Viable alpha threshold
IC > 0.05; full-pipeline latency targetp95 < 450ms.
Constraints & limitations.
Hybrid architecture is superior for most institutional workflows, but it is not free.
- Engineering complexity. Five layers mean more interfaces, observability, versioning, and failure surfaces. Integration quality becomes a first-order research problem.
- Coordination latency. A full pipeline at p95 ~450ms is far slower than a pure-classical signal engine at ~50ms — fine for portfolio construction, wrong for ultra-low-latency HFT.
- Time-series foundation models are still immature. Public models remain undertrained on financial structure; even strong internal models may not generalise across asset classes, horizons, or liquidity regimes.
- Human-in-loop adds delay exactly when situations are novel. The correct trade-off for institutional safety, but not a universal replacement for every architecture.
- Vendor concentration. NVIDIA NIM introduces recurring cost (~$4,500/GPU/year under these assumptions) and platform dependency — be honest about it before romanticising AI modularity.
- IC ~0.08 is a Nyquist internal estimate, not a universal constant. The architecture generalises more reliably than any specific number.
Key takeaways.
- Pure-AI stacks fail on five fronts: numerical hallucination, temporal leakage, low-SNR weakness, regime blindness, governance failure.
- Pure-classical stacks fail on five equally serious fronts: semantic blindness, alt-data scaling limits, slow iteration, model staleness, human monitoring bottlenecks.
- The right architecture is not ideological — it is allocative. Give each layer the task class it can actually solve under production constraints.
- The single most important rule is operational: LLMs should never perform arithmetic that can affect a trade, a risk metric, a margin figure, or an execution decision.
- A robust institutional stack in 2026 has five layers: semantic LLMs, specialised time-series models, classical optimisation, symbolic compute, human governance.
- Firms deploying hybrid stacks see 2–3× faster strategy development than pure-classical peers, and far fewer look-ahead-bias incidents than pure-AI deployments.
The collapse from Sharpe 1.4 to 0.6 in eight weeks was not bad luck and not a regime shift. It was a category error: deterministic numerical work handed to a probabilistic language system, look-ahead bias engineered in by design, no governance layer to catch the decay. The fix is not more AI, and it is not less. It is correct decomposition — five layers, one operational invariant.