LLMs versus Conventional ML
Three families, not two — task-trained models, domain foundation models, and general LLMs, and which one belongs where in a market system
The instinct is to stage a duel between LLMs and ‘traditional’ machine learning and crown a winner. That framing is already out of date: by 2026 there are three distinct families, and the useful question is not which is best but which belongs where — a discrimination a market system depends on, and one the efficient market puts a hard ceiling on.
The instinct is to frame this as a duel — large language models versus “traditional” machine learning — and pick a winner. That framing was always a little lazy, and by 2026 it is simply wrong: there are now three distinct families of model a market system might use, not two, and the interesting question is not which is best in the abstract but which belongs where. Getting that allocation right is most of the engineering, and getting it wrong — reaching for a general LLM to do a job a trained model does better, cheaper and more honestly — is one of the most common and expensive mistakes in applied AI. This entry draws the three families apart, is honest about a frontier that has genuinely shifted, and sets the ceiling that the rest of this site’s evidence imposes on all of them.
Family one — task-trained models (the whole ML library)
The classical picture: a model with weights fit to your data, for one task. A logistic regression, a gradient-boosted tree, a small neural net — every entry in this site’s Machine Learning Library is this family. It is trained once on your labelled examples and then answers that one question, fast. Its virtues are exactly the ones a deployed risk system cares about: it is cheap and low-latency (microseconds, effectively free per prediction), deterministic (same input, same output), it can be calibrated so its probabilities mean what they say, it is auditable via SHAP or its own coefficients, and it cannot hallucinate — it only knows the features you gave it. For a specific prediction on structured numeric data, with training data to hand, this is the default, and it is a very good one.
Family two — domain foundation models (the 2026 shift)
The line between “trained on your task” and “pretrained and prompted” — long the clean divide between ML and LLMs — has been blurred by a third family that is neither. Domain foundation models are pretrained on many datasets and then used zero-shot or in-context on numeric data: for tabular data the TabPFN / TabICL / TabFM lineage, for time series models such as Chronos, TimesFM and TimeGPT. You hand a tabular foundation model the labelled rows as context and it predicts the held-out rows in a single forward pass — no training loop at all. And the frontier has genuinely moved: on the main public tabular benchmarks, the best single models are now these foundation models, sitting a meaningful margin above the best tuned gradient-boosted tree. What they buy is sample efficiency (strong performance from little data) and zero-shot convenience (skip the training and tuning). What they are emphatically not is a general chat LLM — you would not ask Claude to forecast a return series; you would reach for a time-series foundation model built for exactly that. Keeping this distinction sharp matters, because the loose phrase “just use an LLM for it” almost always means the wrong family.
Family three — general LLMs (Claude, GPT)
A large language model is pretrained on internet-scale text and steered not by training but by prompting — frozen weights, behaviour shaped by the context you provide. Its defining strength is the one neither other family has: turning unstructured text into structure, and reasoning in natural language over messy, novel inputs. Give it a news article and ask for the event type, the affected tickers and the sentiment, and it will do in one prompt what would otherwise take a bespoke NLP pipeline. Its weaknesses are the mirror image of family one’s strengths: it is expensive and slow (hundreds of milliseconds to seconds, real cost per call), non-deterministic, poorly calibrated out of the box, opaque, and it can hallucinate — inventing a number or a fact with complete fluency. For anything numeric and repetitive, those are disqualifying. For text, they are the price of a capability nothing else offers.
The honest trade-off
| Task-trained model | Domain foundation model | General LLM | |
|---|---|---|---|
| Input | tabular / numeric | tabular / time-series | text (any) |
| Adaptation | trained on your task | zero-shot / in-context | prompted (frozen) |
| Cost & latency | tiny (µs, ≈ free) | moderate | high (100 ms–s, $) |
| Determinism | deterministic | mostly | stochastic |
| Calibration | can be calibrated | reasonable | poor out-of-box |
| Hallucination | none | none | yes |
| Interpretability | SHAP / coefficients | limited | hard |
| Shines at | a specific task, with data | few-shot / scarce data | unstructured text → structure |
The ceiling nobody escapes
Before allocating families to jobs, one result from the rest of this site overrides everything: in a near-efficient market, next-day direction is unpredictable at every level of sophistication — the capstone shows a transformer tying a logistic regression at a coin-flip AUC of 0.50. That ceiling applies to all three families equally. A tabular foundation model does not conjure an edge a gradient-boosted tree could not find; a general LLM asked to predict tomorrow’s return is not secretly an oracle. Foundation models buy sample efficiency and convenience, not alpha, and no amount of pretraining repeals market efficiency. So the choice between families is almost never about predictive power on returns — there is little to be had — and almost always about the engineering properties that decide whether a system is deployable: cost, calibration, determinism and auditability. That reframing is itself the most useful thing this comparison offers.
Which belongs where in the proposal
The architecture allocates each family to what it is actually good at:
- The News & Narrative Agent needs a general LLM — text → structure is its irreplaceable job, and no other family can do it. This is the only place a general LLM belongs in the system.
- Regime, volatility and sizing are numeric prediction, so they use task-trained models by default (calibrated, cheap, auditable — the ML library and RL tier), with domain foundation models a reasonable option where data is scarce.
- The mistake to avoid, explicitly: do not reach for a general chat LLM to predict a number. It is expensive, uncalibrated, non-deterministic theatre, with no edge to show for it — and using one big LLM for everything is precisely the single-opaque-model anti-pattern the multi-agent design exists to reject.
Grounding in MarketLens AI
My MarketLens AI app already respects this split, mostly. It uses a general LLM (Claude) for the text work — interpreting news for the sentiment dashboard, drafting the prose of an equity-research report — which is exactly family three’s domain. Where it would be tempting to go wrong is a numeric signal: if MarketLens added a next-day direction or volatility indicator, the right tool is a trained, calibrated classifier (family one), not a prompt asking the model to emit “72% chance up.” That number would look authoritative and mean nothing — uncalibrated, non-reproducible, and, per the ceiling above, edgeless. The discipline is to let the LLM do the reading and let a small honest model do the counting.
How I would explain it to a supervisor
“I’d resist framing it as LLMs versus traditional ML, because there are three families now. Task-trained models — the whole ML library — are cheap, calibrated, deterministic and auditable, and they’re the default for numeric prediction. Domain foundation models, the tabular and time-series pretrained models that in 2026 actually beat gradient boosting on benchmarks, add zero-shot convenience and sample efficiency but aren’t chat LLMs. And general LLMs are unmatched at turning unstructured text into structure but are expensive, stochastic, uncalibrated and can hallucinate. The key point is that the choice is rarely about predictive power on returns — my own capstone shows the market is near-efficient, so a transformer ties a logistic regression at 0.50, and that ceiling binds every family — it’s about engineering properties. So in the proposal the general LLM does exactly one job, the News Agent’s text-to-structure, and everything numeric uses trained or domain-foundation models. The anti-pattern I most want to avoid is asking a chat LLM to predict a number: it produces confident, uncalibrated nonsense, which is the opposite of what a risk system needs.”
The 2026 frontier is represented honestly: tabular foundation models (TabPFN / TabICL / TabFM) and time-series foundation models (Chronos, TimesFM, TimeGPT) now rival or exceed tuned gradient-boosted trees on public benchmarks — but they are domain foundation models, distinct from general chat LLMs, and the near-efficiency of returns (established in the capstone) caps the predictive edge of all three families. Claims about MarketLens AI reflect its current design (an LLM used for text tasks). Model-family capabilities move quickly; the engineering trade-offs — cost, calibration, determinism, auditability — are the stable part.