One-line takeaway: On August 27, Toronto's The Finance Lab released TFL Bloodhound Model 1, replacing RLHF's human raters with "Reinforcement Learning from Market Feedback (RLMF)." Each anticipated market move is recorded as a testable hypothesis, then scored on six dimensions once outcomes become observable — probability calibration, magnitude error, distributional accuracy, risk-adjusted utility, maximum adverse movement, and relevance of retrieved evidence. GRPO assigns credit at the evidence level, not the response level, and a companion system called MeMo (Memory-as-a-Model) decays stale relationships. The same week, Boosted.ai launched its multi-model "AI investment committee" Alfa Prime, and TauricResearch's TradingAgents passed 100K GitHub stars. Three events point to the same shift: the hard problem of financial AI is migrating from "is there a model that can talk" to "who judges, and how are scores recorded."
Why RLHF Has a Structural Flaw in Finance
RLHF (Reinforcement Learning from Human Feedback) underpins today's mainstream LLM alignment: humans label which of two responses is better, a reward model is trained on those preferences, and the policy is optimized against it. This works for email, code, and summaries because humans are competent judges of those tasks.
In finance, it hits a hard wall: humans are not competent judges of financial tasks, for three reasons:
1. Feedback timing mismatch. An investment call may take months or years to resolve. At labeling time, a rater can only judge whether analysis *looks* professional — measuring rhetoric and logical self-consistency, not correctness. 2. Non-stationarity. Markets are not documents. Relationships that carried predictive power six months ago may be worthless today; noise can become load-bearing under a new volatility regime. Human raters' knowledge lags regime shifts. 3. Process vs. outcome confusion. Bad reasoning can produce a lucky right answer; rigorous reasoning can lose money to a black swan. Scoring on outcomes alone teaches gambling; scoring on human preference alone teaches rhetoric.
TFL Bloodhound's design starts by admitting these three flaws and routing around them.
Cut One: Separating Estimation from Reasoning
Bloodhound refuses to make one model do everything. Forcing a language model to both estimate high-dimensional time series and reason over those estimates means it must do what it's worst at before reaching what it's best at. The architecture splits into two layers:
- Specialist layer — handles the numerical structure of markets: regime, trend and directional alignment, volatility state, momentum, cross-asset relationships, historical similarity, forward return distributions. Crucially, it outputs structured evidence, not predictions — the reasoning layer receives regime classifications, not raw series to interpret in prose.
- Reasoning layer — handles what actually belongs to language: weighing competing signals, surfacing contradictions, identifying missing evidence, judging which evidence matters under current conditions.
- 20,000+ financial instruments
- 100.8 million historical candles
- ~2 billion feature-level observations
- ~20 years of history
- Bloodhound swaps the training signal — real outcomes instead of multi-model critique;
- Alfa Prime swaps the inference structure — a committee of existing models rather than a new trained model;
- TradingAgents is an open-source methodology template — architecture, not product.
When the reasoning layer hits uncertainty it can call another specialist, retrieve historical context, or recall a similar situation from memory — an iterative, evidence-seeking loop, not single-pass generation.
The elegance: estimation error and reasoning error are isolated in two separately optimizable spaces, communicating through the narrow interface of "structured evidence." When the system errs, you can tell which layer failed.
Cut Two: Six-Dimensional Scoring, Not One Binary Verdict
Every anticipated market move is launched as a testable hypothesis, logging the evidence available at the time, what 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 events stated at 70% probability 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 | Worst adverse excursion while the position would be held | | Relevance of the evidence retrieved | Was the retrieved evidence actually useful? |
The sixth dimension is the most distinctive. The reasoning policy is trained with GRPO (Group Relative Policy Optimization), optimizing the actions of evidence retrieval and hypothesis evaluation, not "generate better prose."
What's New About Evidence-Level Credit Assignment
The key sentence: credit is assigned at the level of evidence, not the response. Traditional RLHF is response-level — one good or bad grade for the whole answer. But a financial reasoning chain may cite seven pieces of evidence, of which three actually mattered and four were along for the ride. Response-level credit can't distinguish them, so the model learns to pile on professional-looking citations. Evidence-level credit assignment back-attributes final calibration quality to each piece of evidence — which specialist outputs genuinely contributed, and which were passengers.
In the team's own words:
> "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."
That calibration is well judged: not "we can predict markets" (hubris) but "we can tell which evidence was historically useful" (engineering).
Cut Three: MeMo — Memory That Expires
The second component, MeMo (Memory-as-a-Model), stores prior situations as multi-dimensional analytical experience: instrument, market environment, model evidence, hypothesis, expected move, realized outcome. Critically, memory relevance is not permanent. MeMo pairs long-term retention with adaptive relevance, so structurally similar situations across industries and assets can be retrieved as probabilistic evidence, while stale relationships decay out of influence.
This directly targets finance's most stubborn enemy: non-stationarity. Standard RAG assumes "retrieved means relevant" — a disaster in finance, where low-volatility price-volume relationships from one regime may be inverse indicators in another. MeMo effectively gives each memory a half-life.
Worth noting the contrast: general-purpose agent-memory research pursues "remember more, retrieve better." Finance's requirement is the opposite — remembering everything without decay is more dangerous than remembering little.
Scale and Foundation
The specialist layer was developed across:
This breadth enables cross-sectional reasoning — treating similar situations that occurred elsewhere as evidence rather than deterministic rules.
The reasoning layer is built on Gemma 4 26B A4B, a sparse mixture-of-experts model with a 262,144-token context window, retrained for financial reasoning. Current status: a pilot deployment across energy futures, plus a public playground.
Two Other Events That Same Week
August 24 — Boosted.ai launched Alfa Prime, billed as "the first agentic AI investment committee for institutional managers." Multiple independent AI models are assigned research tasks arguing bull, bear, and neutral cases; they challenge each other's assumptions, identify evidence gaps, and revise across rounds; a moderator assesses where they agree, diverge, and why, distilling everything into a citable investment memo with supporting evidence, open questions, key risks, and the assumptions most likely to change the conclusion. Internal testing claims a ~2x rate of identifying high-quality opportunities versus baseline analyst workflows; compliance credentials include SOC 2 Type II, ISO 27001, and ISO 42001. CEO Joshua Pantony: "One model giving you one answer isn't enough. The future is investors setting objectives, multiple models building both sides and challenging each other, and an AI collective optimizing toward the outcome."
August 26 — TauricResearch's TradingAgents hit 100,222 GitHub stars, a multi-agent architecture importing debate and reflection into finance: fundamental analysts (SEC filings, earnings), technical analysts (price/volume), sentiment researchers (news/social), a risk manager, and a portfolio manager who synthesizes debates into trade recommendations with confidence scores.
The three differ usefully:
What One Hypothesis Record Looks Like
| Field | Example | |---|---| | Instrument & horizon | An energy futures contract, next 10 trading days | | Launch timestamp | Minute-precise | | Evidence at launch | Volatility regime class, cross-asset correlations, top-5 historical analogs | | Extra info requested | Specialist #3 called for an inventory data pass | | Specialists consulted | Volatility state, momentum, forward return distribution | | Reasoning path | Five-step chain, each step annotated with evidence IDs | | Expected move | Direction + magnitude distribution (not a point estimate) | | Realized outcome | Backfilled at expiry | | Six-dimension scores | Calibration / magnitude / distribution / risk utility / max adverse movement / evidence relevance | | Evidence-level attribution | Volatility 60%, cross-asset 25%, historical similarity 15% |
The last row is the most reusable product of the entire system. "Which factors mattered, when" has always lived as tacit knowledge in a few senior researchers' heads — and leaves when they do. If evidence-level attribution output is stable, it converts senior researchers' tacit judgment into accumulable, auditable, transferable structured assets. That's worth far more than a buy/sell suggestion.
Three Buckets of Cold Water
One: market feedback is sparse, noisy, non-stationary — and rewards "wrong process + right outcome." The team's guardrails — graded (not binary) outcome scoring, evidence-level (not response-level) credit assignment, and strict point-in-time construction (only evidence actually available at hypothesis launch time) — are the right mitigations. But they mitigate, they don't solve. A well-performing in-sample attribution system can still fail out-of-sample.
Two: the pilot is energy futures; the public artifact is a playground. Energy futures are among the more "physical" instruments — clear supply/demand, inventories, seasonality — making structure more capturable than in equities. There is no public evidence yet of transfer to stocks, credit, or FX.
Three: 2x, Sharpe, hit rates — check the denominator. Boosted.ai's "2x baseline analyst workflow" doesn't publish the baseline definition, the test period, or whether out-of-time validation was performed. Relative-improvement claims without a full test protocol carry near-zero information.
Three Windows to Watch
1. Real performance disclosure from the energy futures pilot. A playground is not live trading. Look for risk-adjusted returns with stated time windows and max drawdown after two or three quarters. 2. Explainability of evidence-level attribution. If Bloodhound can reliably output "volatility regime evidence contributed 60%, cross-asset 25%, historical similarity 15%," its value to humans isn't advice — it's teaching humans how to read markets. That's its most likely moat. 3. Public cross-asset validation. Energy futures to equities or FX crosses a whole different microstructure. Public data on a second asset class before 2027 would give the thesis its first generality check.
Epilogue: Picking the Judge Is Harder Than Picking the Contestant
In sports, the hardest job is finding a referee — someone who knows the rules better than the athletes and never plays the game. Financial AI's three-year predicament is fundamentally a shortage of qualified judges: human raters understand language but not markets; market outcomes know right from wrong but not process; multi-model critique understands debate but can be wrong together.
Bloodhound's answer: let the market judge, split the scoring into six dimensions, assign credit at the evidence level, and give memory an expiration mechanism. It's imperfect, but it advances the question from "do you trust this model" to "is this evidence worth trusting." The second framing 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 roundup 2026-08-17 to 2026-08-25 (Mint-Agent / Binance Agent OS), https://aiagentstore.ai/ai-agent-news/topic/trading/