Let the Market Be the Judge: Swapping RLHF's Human Raters for Market Settlements

> One-line summary: On August 27, Toronto-based The Finance Lab released TFL Bloodhound Model 1, replacing RLHF's human evaluators with Reinforcement Learning from Market Feedback (RLMF). Every expected market move is logged as a testable hypothesis; once outcomes become observable, the model is scored on six dimensions—probability calibration, magnitude error, distributional accuracy, risk-adjusted utility, maximum adverse movement, and evidence relevance—with GRPO assigning credit at the evidence level rather than the response level. A companion component, MeMo (Memory-as-a-Model), lets stale relationships decay automatically. The same week, Boosted.ai launched its multi-model "AI investment committee" Alfa Prime, and TauricResearch's TradingAgents crossed 100,000 stars. All three point to the same shift: the hard problem in financial AI is moving from "is there a model that can talk" to "who grades the answers, and how are the scores kept."

RLHF has a built-in flaw for finance

RLHF (Reinforcement Learning from Human Feedback) underpins today's mainstream LLMs: humans compare two responses, a reward model learns those preferences, and the policy is optimized against it. This works because humans are competent judges for tasks like writing emails or code. In finance, humans are not competent judges, for three reasons:

1. Feedback latency mismatch. An investment call may take months or years to resolve. A human annotator at answer-time can only rate how professional the analysis *looks*—style and internal coherence, not correctness. 2. Non-stationarity. Markets are not documents. A relationship predictive six months ago may be worthless today; a previously noisy one may suddenly matter under a new volatility regime. Annotators' knowledge lags regime switches. 3. Process vs. outcome confusion. A flawed process can stumble into a right answer; a rigorous process can lose money to a black swan. Score only outcomes and the model learns gambling; score only human preference and the model learns rhetoric.

Bloodhound's design starts by acknowledging these three flaws and routing around them.

Cut 1: Separate estimation from reasoning

Bloodhound does not ask one model to do everything. Forcing a language model to both estimate high-dimensional time series and reason over those estimates means it must first do what it is worst at before reaching what it is best at. So the architecture splits into two layers:

• Specialist layer: quantitative models covering market regime, trend and directional alignment, volatility state, momentum, cross-asset relationships, historical similarity, and forward return distributions. Crucially, they output structured evidence, not predictions—e.g., "currently in this volatility regime"—not raw series for the reasoning layer to interpret in prose.
• Reasoning layer: the genuinely linguistic work—weighing competing signals, surfacing contradictions, spotting what is missing, judging which evidence matters under current conditions. On uncertainty, it can call another specialist, retrieve historical context, or recall a similar episode from memory. The loop is iterative and actively evidence-seeking, not single-pass generation.

The elegance: estimation error and reasoning error are isolated into two separately optimizable spaces, communicating through the narrow interface of structured evidence—classic separation of concerns.

Cut 2: Six-dimensional scoring, not right-or-wrong

Each expected market move is launched as a testable hypothesis, recording the evidence available at the time, which information the model requested, which specialists it consulted, and its reasoning path. When outcomes become observable, reward is computed on six dimensions:

| Dimension | What it measures | |---|---| | Probability calibration | Did 70%-probability events actually occur ~70% of the time | | Magnitude error | Deviation between expected and realized magnitude | | Distributional accuracy | Does the predicted return distribution match the realized one | | Risk-adjusted utility | Utility after risk adjustment | | Maximum adverse movement | Deepest drawdown while the position was live | | Relevance of retrieved evidence | Was the retrieved evidence actually useful |

The reasoning policy is trained with GRPO (Group Relative Policy Optimization), optimizing evidence retrieval and hypothesis evaluation rather than "generate nicer prose."

What's new about evidence-level credit assignment

Credit is assigned at the level of evidence, not the response. RLHF is response-level: the whole answer is rewarded or penalized together. But a financial reasoning chain may cite seven pieces of evidence, three of which actually mattered. Response-level credit can't tell them apart, so the model learns to pile on professional-looking citations. Evidence-level assignment back-propagates final calibration quality onto each piece of evidence—a harder but far more valuable reward signal. As the team puts it:

> "This is not a statement about whether markets are predictable. It's a statement about whether which evidence mattered is answerable after the fact—and that is a usable training signal."

Cut 3: MeMo — memory that expires

MeMo (Memory-as-a-Model) stores prior episodes as multi-dimensional analytical experience: instrument, market environment, model evidence, hypothesis, expected move, realized result. The key: memory relevance is not permanent. MeMo pairs long-term retention with adaptive relevance, so structurally similar episodes across sectors and assets can be retrieved as probabilistic evidence, while stale relationships decay out of influence. This directly targets finance's oldest enemy—non-stationarity. Standard RAG assumes whatever is retrieved is relevant, which is catastrophic in finance. MeMo effectively gives every memory a half-life. Notably, this runs *against* the general agent-memory trend of "remember more, retrieve better": in finance, remembering a lot without decay is more dangerous than remembering little.

Scale and base model

• 20,000+ financial instruments
• 100.8 million historical candles
• ~2 billion feature-level observations
• ~20 years of history

This breadth enables cross-sectional inference—treating similar episodes elsewhere as evidence, not deterministic rules. The reasoning layer is built on Gemma 4 26B A4B, a sparse MoE model with a 262,144-token context window, further trained for financial reasoning. Current status: pilot deployment across energy futures, with a public playground.

Two other events that same week: judge and jury

August 24 — Boosted.ai launched Alfa Prime, billed as the first agentic AI investment committee for institutional managers. Multiple independent models are assigned research tasks arguing bull, bear, and neutral cases; they challenge each other's assumptions, identify evidence gaps, and revise over multiple rounds; a moderator assesses where they agree, disagree, and why—distilled into a citable investment memo with supporting evidence, open questions, key risks, and the assumptions most likely to change the conclusion. Boosted.ai's internal testing claims a ~2x rate of identifying high-quality opportunities versus a baseline analyst workflow; compliance credentials include SOC 2 Type II, ISO 27001, and ISO 42001. CEO Joshua Pantony: "One model giving you one answer isn't enough."

August 26 — TauricResearch's TradingAgents hit 100,222 GitHub stars. Its architecture brings multi-agent debate and reflection to finance: fundamental analysts (SEC filings, earnings), technical analysts (price/volume), sentiment researchers (news, social), a risk manager, and a portfolio manager who synthesizes the debate into trade suggestions with confidence scores.

The three differ in kind:

• Bloodhound changes the training signal—scored by realized market outcomes, not peer debate.
• Alfa Prime changes the inference structure—a committee of existing models, no new training.
• TradingAgents is an open-source methodology template—architecture, not product.

These layers are complementary: training signal bounds capability, inference structure shapes per-decision quality, and the open ecosystem determines how fast methods spread.

What one hypothesis record looks like

| Field | Example | |---|---| | Instrument & horizon | An energy future, next 10 trading days | | Launch timestamp | Minute-precise | | Evidence available at launch | Volatility regime, cross-asset correlations, top-5 historical analogs | | Information requested | Specialist #3 called for inventory data | | Specialists consulted | Volatility state, momentum, forward return distribution | | Reasoning path | Five-step chain, each step tagged with evidence IDs | | Expected move | Direction + magnitude distribution (not point estimate) | | Realized outcome | Backfilled at expiry | | Six-dimension score | Calibration / magnitude / distribution / risk utility / adverse movement / evidence relevance | | Evidence-level attribution | Volatility 60%, cross-asset 25%, historical similarity 15% |

The last row is the most reusable artifact. "Which factors mattered, and when" has always lived as tacit knowledge in a few senior researchers' heads—lost when they leave. If evidence-level attribution works reliably, it converts tacit judgment into structured, auditable, transferable institutional knowledge. That is worth far more than a buy/sell suggestion.

Three buckets of cold water

1. Market feedback is sparse, noisy, non-stationary—and rewards "bad process + lucky outcome." The team's guardrails—graded (not binary) outcome scoring, evidence-level credit, strict point-in-time construction—mitigate but do not solve this. In-sample attribution can still fail out-of-sample. 2. The pilot is energy futures; only a playground is public. Energy futures are among the more "physical" instruments, with clear supply/demand and seasonality. No public evidence yet on transfer to equities, credit, or FX. 3. "2x," Sharpe, win rate—all depend on the denominator. Boosted.ai's "2x baseline" leaves the baseline definition, test period, and out-of-time validation undisclosed. Such claims carry near-zero information without a full test protocol.

Three windows to watch

1. Real performance disclosure from the energy futures pilot—time-stamped risk-adjusted returns and drawdown after a few quarters, not playground results. 2. Interpretable evidence-attribution output—if Bloodhound consistently reports which evidence carried what weight, its moat is "teaching humans to read markets," not just giving advice. 3. Public cross-asset validation—if a second asset class shows public data before 2027, the route's generality gets a first answer.

Epilogue: Picking the judge is harder than picking the athlete

In sports, the hardest part is finding the referee—someone who knows the rules better than the athletes but never steps onto the field. Financial AI's three-year predicament is fundamentally the absence of a competent judge: human annotators understand language but not markets; market outcomes know right from wrong but not process; model-to-model debate knows argument but may be wrong together. Bloodhound's answer: let the market be the judge, but split scoring into six dimensions, assign credit at the evidence level, and give memory an expiry mechanism. It's imperfect, but it moves the question from "do you trust this model" to "is this evidence worth trusting"—and the second question is far more useful.

References

1. ACCESS Newswire, "The Finance Lab Introduces TFL Bloodhound, a Financial Reasoning Model Trained on Market Outcomes Instead of Human Preference," 2026-08-27, https://www.accessnewswire.com/newsroom/en/computers-technology-and-internet/the-finance-lab-introduces-tfl-bloodhound-a-financial-reasoning-m-1213198 2. Business Wire / Morningstar, "Boosted.ai Introduces a Multi-Model AI Investment Committee for Select Institutional Partners," 2026-08-24, https://www.morningstar.com/news/business-wire/20260824772607/boostedai-introduces-a-multi-model-ai-investment-committee-for-select-institutional-partners 3. FF News, "Boosted.ai Launches Alfa Prime: The First Agentic AI Investment Committee for Institutional Managers," 2026-08-24, https://ffnews.com/news/boosted-ai-introduces-a-multi-model-ai-investment-committee-for-select-instituti-a8de6153 4. Smartotics, "AI Daily Report — 2026-08-26" (TradingAgents 100,222 stars entry), https://smartotics.com/2026/08/2026-08-26-ai-daily-2026-08-26.html 5. AI Agent Store, "Weekly signal," agentic trading overview 2026-08-17 to 2026-08-25 (Mint-Agent / Binance Agent OS), https://aiagentstore.ai/ai-agent-news/topic/trading/