Spokes.wiki Search About

Synthesis — Quant Trading

The evolving thesis. Spun out of the hub _inbox financial-ml cluster on 2026-06-29, once a third coherent source landed: a financial-market foundation model (kronos-financial-foundation-model, parked first), a reinforcement-learning trading framework (tensortrade, parked second), and an event-driven trading bot (banbot, the trigger that completed the cluster).

Current thesis

A trading system is a stack of three layers, and the founding trio lands one source on each:

  1. Predict — Kronos forecasts future OHLCV bars from price history, borrowing the text-LLM tokenize-then-autoregress recipe and retargeting it to the noisy financial modality.
  2. Decide (strategy) — turn information into a position, either as a learned policy (TensorTrade's RL agents) or a hand-coded rule set (the strategies banbot runs).
  3. Execute — banbot's event-driven engine places and manages the orders against live exchanges, multi-symbol and multi-timeframe.

The through-line binding all three is trustworthy backtesting: a strategy is only as good as the simulation that validated it, and the cardinal failure is lookahead bias — letting a decision use information it couldn't have had in real time. The shared structural answer is the event-driven design — a forward-only event stream makes peeking at the future impossible — which is exactly why it also enables backtest/live parity (one codebase, two modes). Kronos's "observe only the past," TensorTrade's observer discipline, and banbot's event loop are three faces of the same constraint.

The shape of the field (early)

Open questions

Growth edges

Ranked; each names the kind of source that would close it (see ../QUALITY.md → Growth edges).

  1. A source that reports its N. The evidence question is now specific: no strategy or engine here discloses how many configurations were searched. — needs: any T1/T2 backtest or engine doc that states its trial count. Half closed 2026-08-08 (research pass): deflated-sharpe-ratio (Bailey & López de Prado, JPM 40(5), T1) supplies the correction the edge named, and sharpens why the omission matters — the trial count is an input to the statistic, so a backtest published without N cannot be deflated at all and is uninterpretable rather than merely thin. The half still open is the original ask: a published strategy or engine that states its own N. Advanced again 2026-08-14 (research pass): thousands-of-alpha-tests (Giglio/Liao/Xiu, RFS 34(7) 2021, T1) states its N in the open — 1,761 hedge funds — and runs an FDR-controlled test across all of them, finding 19% with real alpha (70bp/month) that holds out of sample. But it is N as fund count (how many existing strategies were compared), not N as configuration count (how many parameter variations of one strategy were searched before publishing the winner — the strategy-optimization case this edge originally meant). Same statistical problem, different object. Residual, narrowed: a source that reports its own single-strategy search count remains unfound; a source reporting a multi-strategy trial count no longer does. Half closed a third way 2026-09-16 (research pass): all-that-glitters (Wiecki, Campbell, Lent & Stauth, J. Investing 25(3) 2016, T1) measures the thing the edge is about instead of naming it — across 888 Quantopian strategies with real out-of-sample records, in-sample Sharpe predicts out-of-sample Sharpe at R² = 0.02, and Sharpe shortfall rises with the log of how many days a strategy was backtested during development. That is search intensity as a proxy, not a configuration count, so the single-strategy N is still unreported anywhere here. Residue unchanged.
  2. The execution layer has no independent measurement. Slippage, fill quality and latency are asserted by the engines themselves. — needs: T1/T2 market-microstructure or execution-quality study.
  3. Nothing on risk or portfolio construction. The corpus stops at "does the strategy make money" and never reaches position sizing, drawdown control or capital allocation. — needs: T1 text or paper on portfolio risk management. Half opened 2026-09-09 by cover-universal-portfolios (T1): constant-rebalanced-portfolio and universal-portfolio give the spoke capital allocation across assets — fixed proportions, rebalancing as a return source, and a rule for choosing the proportions without foreknowledge. It also drags in the tradition the spoke had only heard of through this paper's citations (Markowitz 1952, Sharpe 1963, Kelly 1956), none of which is paged. Residual: risk proper — position sizing against drawdown, capital at risk, and the mean–variance frame itself. Cover's construction has no risk term at all; it maximizes growth rate and says nothing about the path.

Coverage edges (added 2026-08-08, at the curator's request for a wider backlog). These widen what the spoke covers instead of answering an open question above; one ordinary solid source closes any.

  1. Derivatives. Half closed 2026-08-11, and the remaining half mostly closed 2026-08-16 by a curator-directed research pass. The original paper came first: black-scholes (T1, JPE 81(3), 1973) with option-pricing as the concept node — the hedge argument, the formula, delta as hedge ratio, implied volatility as the inversion the market actually quotes. The 2026-08-16 pass asked for exactly what the edge named, a source measuring a hedging practice rather than listing the quantities, and got it: pervasive-impact-of-option-trading (T1) defines gamma and measures what rebalancing against it does to the underlying, gamma-fragility (T2) repeats the measurement from public data, and information-in-option-volume (T1) supplies the flow side (options-greeks, dealer-gamma-exposure, unusual-options-activity). Residual, narrowed: vega, theta and rho. Delta and gamma are now defined and measured here; the volatility, time and rate sensitivities are still named only as a feature bullet in staskh-trading-skills — needs: a derivatives text chapter, which is the one form of source this pass did not find at T1/T2 and did not settle for below it.
  2. Market microstructure. market-making describes a role inside a venue the corpus never describes: the limit order book, order types, queue position, Reg NMS. — needs: a microstructure text or an exchange's own matching-engine documentation.
  3. The anomaly literature. backtest-overfitting and pseudo-mathematics-financial-charlatanism are critiques of a body of published factor research — Fama-French, momentum — that the spoke does not hold. — needs: the original factor papers, T1.
  4. Data and its defects. Survivorship bias, point-in-time fundamentals and corporate-action adjustment silently decide every backtest here. — needs: a paper quantifying one of these biases, or a data vendor's methodology.

The marketing layer arrives, as a specimen

ai-trading-bot-platforms-2026 is the first source here about the retail platform market — products that sell automated trading to people who will never write an event loop — and it is T4: a ranked roundup whose byline is "Marketing/IT Dept.", in which the top pick carries every link in the piece with a tracking parameter attached and the other four platforms carry none, with no disclosure and nothing tested.

Ingested rather than refused, because two things in it are worth keeping.

The category map. Five products named and described (3Commas, Pionex, Cryptohopper, Trade Ideas, SaintQuant), all unverified, plus a control axis that is genuinely the right way to cut the category: managed → configurable → exchange-native → signal-only (trading-bot-platform). It maps onto this wiki's own predict/decide/execute stack — a managed platform sells all three as one product, a configurable one sells execution plus a toolkit, a signal-only one sells prediction and leaves the rest with you. banbot and tensortrade are the open-source form of the middle.

The vocabulary gap, which is the actual finding. Nothing in the article mentions Sharpe ratios, out-of-sample testing, transaction costs, slippage or capacity. Not hedged, not glossed — absent. Set that beside pseudo-mathematics-financial-charlatanism, which shows that seven configurations over two years manufacture an in-sample Sharpe of 1 from a strategy with no edge, and the shape of the retail market becomes legible: the things that decide whether an automated strategy works are not the things it is sold on, and a managed platform's risk rules are precisely the part a buyer cannot inspect.

That extends the corpus's standing thread rather than opening a new one. The tension section below already holds "nothing in this corpus can tell you whether a reported edge is real" — this source is the case where nobody even reports one, and it still ranks five products. Its most accurate sentence is its disclaimer: "No bot guarantees profit."

One order flow, two mechanisms (options, added 2026-08-16)

A research pass at the curator's request — "combining options Greeks and unusual options activity" — added three papers and moved the derivatives corner from a formula to a mechanism.

The finding is that option order flow gets read two ways, and the two readings compete for the same observation. information-in-option-volume (Pan & Poteshman, RFS 2006, T1) reads it as direction: a put-call ratio built from volume bought to open new positions predicts the next day's risk-adjusted return at −53bp per unit, over 1% a week in the extreme quintiles, strongest in far-out-of-the-money contracts expiring within 30 days. pervasive-impact-of-option-trading (Pearson, Poteshman & White; published as Ni et al., RFS 2021, T1) reads the same market as position: the net gamma of investors who hedge explains about 12% of an optioned stock's daily absolute return, whether or not anyone is informed. gamma-fragility (Barbon & Buraschi, T2) reproduces the second result from public data and inherits an assumption doing so.

Three things follow, and they are the spoke's new options through-line (unusual-options-activity, dealer-gamma-exposure, options-greeks).

The two mechanisms peak on the same contract. Short-dated and far OTM is where Pan & Poteshman find the most information and where a dealer picks up the most gamma per dollar. So an unusual call-buying print followed by a rising stock is equally consistent with the buyer knowing something and with nobody knowing anything while the hedge does the work. The flow alone cannot separate them, which means "Greeks plus unusual activity" is not two confirmations of one signal — it is one observation with two mechanisms attached.

The evidence gap here is a data-access gap, not a rigour gap. This is a new shape for this spoke. Everywhere else, the corpus's problem is that the author ran the backtest (backtest-overfitting, pseudo-mathematics-financial-charlatanism). Here the statistics are clean and the input is private: the predictive signal is built from a chicago-board-options-exchange field that says whether a trade opened or closed a position, and Pan & Poteshman show that the version an outsider can reconstruct predicts for a day or two, reverses, and contributes nothing once the private signal is in the same regression. gamma-fragility hits the same wall from the other side — holder identity is not public, so it assumes the dealer sign and cites Pan & Poteshman to pick which assumption. Every retail unusual-activity screener and gamma-exposure dashboard is running on the data the research says is the part that does not work.

Nothing here is a strategy, and all three say so. Pan & Poteshman frame their result as price discovery, not an edge; the hedging papers measure a channel and never trade it; no source tests either net of costs. That is the same distance between "a real effect" and "a tradeable one" that option-pricing already recorded for black-scholes's own 1970s deviations.

The corpus's oldest source is the only one that reports where it fails

black-scholes arrives as coverage — the spoke could not describe most of what professional desks trade — and what it actually supplies is a standard of conduct the rest of the corpus does not meet.

It prices without a backtest. Everything else here settles "does this work" by replaying history in a simulator the author wrote (backtesting). Black and Scholes derive a value from a no-arbitrage condition: a continuously adjusted hedged position carries no risk, a riskless position must earn the short rate, and one formula satisfies that constraint. No history is consulted. Options are, so far, the only instrument in this wiki where a claim about value can be checked against theory instead of against a simulation, which is what option-pricing adds to the predict→decide→execute stack rather than merely beside it.

It states its assumptions as a list. Seven ideal conditions, numbered, at the point of use — constant rate, lognormal prices with constant variance, no dividends, European exercise, no transaction costs, unlimited borrowing, free short selling. Every one is false somewhere, and saying so is what lets a reader decide whether the formula applies to their instrument. No backtest in this corpus ships that list. fincept-terminal ships factor discovery, ML and RL with no statement of conditions at all.

And it reports a negative result about itself. The authors' own empirical test found market prices deviating systematically from their formula — buyers consistently overpaying, the gap widest on low-risk stocks — and then closed the loop: given transaction costs, the deviation "does not imply profit opportunities for a speculator." They found their model wrong in a specific direction, and they published that the error was not tradeable.

Fifty-three years later this wiki holds superior-skills validating a strategy on six trades at a 100% win rate, and ai-trading-bot-platforms-2026 ranking five products with no mention of Sharpe, out-of-sample testing, costs or capacity. The corpus's standing worry has been that performance numbers come from simulators their authors wrote. The sharper version, visible only now that there is something to compare against: not one modern source here reports a negative finding about its own method. deflated-sharpe-ratio supplies the statistic that would expose the problem and chan-algorithmic-trading argues for the model class that resists it; this is the older and simpler discipline both are reaching for — say what you assumed, then say where it broke.

One honest caveat about the comparison. A closed-form pricing result and a machine-learning trading strategy are not the same kind of object, and a paper in the JPE answers to reviewers that a GitHub README does not. The point is not that the repositories should have derived a theorem. It is that stating conditions and publishing a failure are cheap, and nobody does them.

A guarantee instead of a backtest (added 2026-09-09)

Cover's "Universal Portfolios" (Mathematical Finance 1(1), 1991, T1) is the first source here whose central claim is proved rather than measured, and it changes what the corpus can say about evidence.

Every other performance figure in this spoke is the output of a simulator: an engine replaying klines, a model scored on a test split, a paper reporting its own backtest. The standing complaint on this page is that such a figure is uninterpretable without the trial count N, and nobody reports one. Cover does not report N either. He doesn't generate one. The universal-portfolio mixes over the entire strategy class instead of selecting from it, weighting each constant-rebalanced-portfolio by the wealth it actually accumulated, and the guarantee is a worst-case bound over arbitrary price sequences: (1/n)ln Ŝₙ − (1/n)ln S*ₙ → 0 for every bounded sequence, with a bound that holds at every finite n. There is no sample to overfit to, and the premise is stated as a refusal — no distribution, no stationarity, crashes allowed by construction.

Three things follow for this corpus.

The evidence question has a fourth answer. It has been asked as "who is the independent referee?" and answered three ways: nobody (T3 self-reports), a structured evaluation toolkit (trademaster), a correction statistic that needs N (deflated-sharpe-ratio). A theorem is the fourth: it needs no referee because it makes no empirical claim. The cost is what it promises: not "this makes money," but "this loses almost nothing against the best fixed-proportion strategy selectable with foreknowledge." If that benchmark loses, so does the algorithm.

Allocation is a layer the predict → decide → execute stack never named. How much of the capital each position gets is a decision every system here makes implicitly and none of them states. Cover's result is about that decision alone, and it lands on the spoke's third growth edge (risk and portfolio construction) — narrowing it, since position sizing against drawdown is still absent.

And the paper's own examples are exactly the evidence this corpus distrusts. §8 demonstrates the algorithm on stock pairs "chosen for their volatility" over one 22-year window ending in 1985, with no count of how many pairs were examined, trading costs excluded outright, and integrals quantized to a 21-point grid. The theorem carries the guarantee; the tables carry none of it. That the strongest source in the spoke on assumption-free reasoning illustrates itself with a hand-picked pair is not a flaw in the proof. It is the clearest sign yet that demonstration and evidence are different activities, and that this corpus has been reading one as the other.

The slope was measured in 2016 (added 2026-09-16)

all-that-glitters is the empirical companion to the two theory papers this spoke has leaned on. Where deflated-sharpe-ratio says a reported Sharpe must be shrunk by the expected maximum under N trials, Quantopian's authors could watch it happen: 888 strategies whose code timestamps fix the in-sample boundary, in-sample Sharpe explaining 2% of out-of-sample variance, and the shortfall between the two growing with how long the author kept testing. Two things follow for the corpus. First, the practitioner sources here that publish a backtest Sharpe with no trial count are not just uninterpretable in principle; the one cohort that was measured says that number carries almost no information on its own. Second, the same paper finds what does carry over — volatility (R² 0.67) and drawdown (0.34) — which is the argument for reading a backtest's risk profile and ignoring its return, and a small vindication of ernest-chan's preference for statistics over the headline figure. The cohort is one platform's users over five years, and the authors sold allocation on the strength of it; the size of the effect should be held loosely, its direction less so.

Contradictions / tensions

Three of the four below turned out to be the same argument seen from different layers, so read them as one thread: nothing in this corpus can tell you whether a reported edge is real. The first names the methodological fork (does simplicity or validation protect you?), the second shows what the fork costs once an LLM can place the order, and the third shows the most prominent artifact in the spoke bundling every layer and reporting discipline for none.

Complexity as the disease, not the cure

The corpus's first genuine internal disagreement, and it is about method rather than a fact.

Everything in this spoke until now bets on more model: kronos-financial-foundation-model is a pretrained transformer over OHLCV bars, tensortrade and trademaster learn policies with deep RL, supervised-learning-trading fits gradient-boosted classifiers to manufactured labels, multithreaded-ml-trading-robot runs an ML ensemble, and strategy-optimization exists to search parameter spaces hard. The shared assumption is that capacity plus a validation protocol beats simplicity.

chan-algorithmic-trading states the opposite as its organizing principle: simple, linear strategies as an antidote to overfitting and data-snooping, with explicit scorn for "the usual 'data-mining' approach" of throwing indicators at a series to see which ones profit. On this view the guard against overfitting is the model class, not the holdout — you cannot fit a two-parameter spread rule to noise the way you can fit a boosted ensemble to it.

Nothing in the corpus adjudicates this, and the two sides are not evenly matched on evidence: Chan is a practitioner asserting a preference in a preface, and the ML sources are self-reported backtests. But it is a real fork, and the spoke's own evidence problem is what makes it live — if every performance number here comes from a simulator its author wrote, then the argument "complex models overfit and you won't be able to tell" is not a stylistic objection, it is a prediction about exactly this corpus. Note also the date asymmetry: Chan is 2013, before the deep-learning wave the rest of the corpus rides.

The agent became the deployment surface

Not a contradiction — a new axis, and one that changes what the evidence problem costs.

superior-skills packages strategy authoring, backtesting, funding, live deployment, sub-account orchestration, withdrawal and a kill switch as Agent Skills, so the whole loop from idea to filled order runs through a conversation with an LLM (agent-deployed-trading). Every other artifact here is a library a programmer drives. A "managed wallet" option is described as "no-key trading, custodial-style UX."

The spoke's standing worry has been that performance numbers come from simulators their authors wrote. Agent deployment adds a second: the last human check is an approval prompt, and the sibling ../agentic-tooling-wiki has spent two months collecting evidence about what such prompts are worth — including a shipping harness conceding that an approval means a human accepted the action as displayed, not that the behaviour is safe. Those findings concern bad commits. Here a mistake is a filled order: a trade has no diff to inspect, and exit-all closes open positions rather than unwinding closed ones.

The commercial shape is the part to keep. The skills are MIT and installable from five entry points; they are inert without an API key to a paid platform. So an open skills pack works as a customer-acquisition channel — publish the interface, keep the backtester, the data and the execution — and the same party then writes the strategy, runs the backtest validating it, and earns on the deployment. Which is visible in the artifact: the repo's headline "validated" strategy is six trades, 100% win rate, 0% drawdown (backtesting).

That number arrived hours after chan-algorithmic-trading gave this spoke data-snooping as a named worry, and it is the cleanest specimen the corpus is likely to get. It does not settle the complexity-versus-simplicity fork above — the Donchian rule is exactly the simple, linear kind Chan argues for — which is itself the useful correction: a simple model is not a defence against a sample of six. Chan's antidote addresses the model class; nothing addresses the evidence.

The terminal bundles every layer, and reports discipline for none

fincept-terminal is the first source that is not one layer of the stack but all of them at once — a Bloomberg-class terminal wrapping data connectors, a QuantLib pricing/risk suite, factor discovery, an ML/HFT/RL "Quant Lab," a paper-trading engine, and live execution through 16 brokers, behind one UI. Its size (29k★) makes it the most prominent single artifact the spoke holds.

It does not adjudicate the complexity-vs-simplicity fork above; it stacks the deck for one side and stays silent on the cost. Everything Chan warned against — searching many indicators, fitting high-capacity models, mining factors — is a menu item here, and the README describes capabilities without a word on validation: no MinBTL, no held-out protocol, no reported trial count N. So it extends the spoke's central evidence problem rather than easing it. The backtest-overfitting worry was that a searched-over Sharpe hides N; a terminal makes the search a few clicks and still does not surface N. It is superior-skills's "6 trades, 100% win" failure shape on a far larger surface, and the useful synthesis is narrow: tooling that lowers the cost of searching configurations, without lowering the cost of reporting how many, moves the field the wrong way on its own hardest problem.

The other thing it marks is the agent-as-terminal turn. agent-deployed-trading arrived through superior-skills as a strategy library run through an LLM; here it is 37 investor-persona agents (Buffett, Graham, Lynch…) with MCP-tool node workflows sitting inside a full terminal. The deployment surface is no longer a skill pack but an application, which raises the same "the last human check is an approval prompt" concern at product scale. The agent machinery itself is ../agentic-tooling-wiki's subject (logged runner-up); what stays here is that a trade authored by a persona agent still has no diff to inspect.

Two skillpacks, and the incentive question that separates them (added 2026-08-06)

staskh-trading-skills is the second agent-skills pack for trading to reach this spoke, and it arrived precisely as a control on what the first one let us conclude. agent-deployed-trading had read superior-skills as an open pack fronting a paid platform — publish the interface, keep the backtester, the data and the execution. staskh is MIT too, and needs no platform: Yahoo Finance for data, the user's own Interactive Brokers account for the portfolio, one optional paid feed the author does not sell. Nobody earns on the deployment.

The correction is small and worth making anyway: openness says nothing about incentives; what the pack needs a key for does. One case looked like a business model, and it was — its own, not the category's.

The more interesting result is that the two packs fail on opposite sides of the spoke's central problem. Superior ships strategies and validates one on six trades. staskh ships no strategies and no backtester at all — 25 skills that pull quotes, compute Greeks and screen for PMCC setups, with nothing between a scanner hit and a position but the trader's judgement. The first is backtest-overfitting in miniature. The second is not an improvement on it: an evidence problem you cannot have because you never make a claim is the same gap with the paperwork removed, and the agent's fluency about RSI and delta does the persuading that a backtest at least pretended to.

The control boundary moved too, and this is the concrete part. staskh defaults Interactive Brokers to read-only, requiring an explicit switch to read-write before orders can be placed. Better than Superior's custodial "no-key trading," and still not review: it is one decision, made once in advance, that covers every order afterwards. The two properties agent-deployed-trading named survive both designs — a trade has no diff, and the kill switch is downstream of the mistake.

One caveat neither pack's README treats as a design constraint: staskh's data is Yahoo's, ~15 minutes delayed, and it is used to compute Greeks and unusual-options-activity scans — the parts of options trading most sensitive to the current underlying price. Fine for study, not for the execution the same feature list implies.

A third design closes half the gap, and a gamified leaderboard reopens it (added 2026-09-27)

beebots is the spoke's third agent-deployed-trading specimen, and it is not a skill pack at all — the decision model, Jev, gets no tool access, only a closed menu of currently valid moves with a probability attached, and a deterministic code layer can veto, shrink or force that answer on every tick. That is a real design improvement on Superior's and staskh's "one decision made once, in advance" control boundary: the check here is continuous, not a switch flipped at setup.

It does not close the other half. Paper trading by default (simulated equity per bee) sidesteps rather than answers "a trade has no diff to inspect" — it just makes being wrong free. And the project's public leaderboard, the Hive, turns out to be this spoke's cleanest live illustration of backtest-overfitting yet: an unstated, presumably growing count of independently-parameterized bees run the same three underlying strategies against the same market, and the board surfaces whichever one got lucky, with no trial count and no significance correction — pseudo-mathematics-financial-charlatanism's best-of-N mechanism, running live and in public, packaged as a game rather than a claim. The project's own copy calling the Hive "a game, not a signal service" is the most honest framing this spoke has seen of exactly this hazard, which is more than superior-skills's "6 trades, validated" managed.

The standing external tension

The older note still holds: self-reported performance vs. the discipline's well-known overfitting and distribution-shift risks — now reinforced by smolinski-h1-2026-aar, whose +33% vs 11% S&P is a trader's own unaudited number reported second-hand (no drawdown/vol context), the same self-report caution the founding bots earn.

Cross-spoke adjacency

Index — Quant Trading Wiki

Catalog of every page, grouped by schema.org @type. Spine: synthesis (thesis), log.md (history), this file (catalog). Spun out of the hub _inbox financial-ml cluster on 2026-06-29 (3 sources).

DefinedTerm (concepts)

Article (sources)

TechArticle (sources)

SoftwareSourceCode (sources)

Collection (sources)

LearningResource (sources)

Organization (entities)

ScholarlyArticle (sources)

Book (sources)

Person (entities)

Synthesis

Notes

Related pages

Links to

Linked from