Kinetic Alpha

Tools

Every claim on this site is a tool you can drive.

Research that asks you to accept its assumptions is asking for trust. These seven tools exist so you do not have to extend any: each one carries the engine behind a published piece, with the assumptions exposed as inputs. Change the ones you disagree with and see whether the conclusion survives. Several do not, under some settings — that is the useful part, and it is why the tools ship alongside the writing rather than after it.

They fall into three kinds, and the difference is about what you bring rather than how complex they are. Platforms carry a whole domain and expect you to arrive with a position or a question. Scanners are pointed at one trade and expect you to arrive wanting a ranking. Companions sit inside a research piece and expect nothing — they make that paper's math drivable while you read it.

Standalone platforms

Four tools built to be used on their own.

These are not illustrations of a paper — each carries its own reference engine and is comprehensive enough to do real work in. They are also the four that took the longest to build, and the ones the consulting work is built against. Each write-up below says what the tool answers, what it is made of, what it runs on, and what to try first.

Compute · PowerMost complete

Compute × Power workbench

What does a rack of GPUs cost to run, and what actually hedges it?

This is the synthesis tool for the whole compute thread, and the one that carries the most of the practice's own arithmetic. It starts at the physical conversion — GPU-hours to kilowatt-hours to grid megawatts, through PUE and the measured duty factors rather than nameplate TDP — and carries that number all the way out to an instrument you could actually trade. In between it prices the compute spark spread the way a generator prices a heat rate, decides whether a load should be hedged as peak, off-peak or around-the-clock, checks whether the basis is congestion (an FTR problem) or energy (a forward problem), and then puts the resulting structure through ICE and Nodal margin so the carry is visible before the trade is on. The regional grid maps five named data-center corridors to the zone that prices them and the instrument that hedges each — North Virginia into PJM/DOM, Texas into ERCOT North, Central Ohio into PJM/AEP, Phoenix into WECC/Palo Verde, and Atlanta into SOCO, which is bilateral and therefore monitor-only. That last row is the useful one: it is the case where the honest answer is that no listed hedge exists.

What's inside
Market stateCompute factors + spark spreadPeak / off-peak / 24×7 shapeTransmission & FTROptions lab (Black-76, Asian, bullet vs monthlies)Trade builderICE / Nodal marginNG seasonalityIndex dispersion (SD vs OCPI)Compute heat mapConvergenceIndex decompositionTake-or-pay vs spotData & method
Runs on

Measured H100 power traces, ERCOT and PJM hourly day-ahead settlements, published index prints and venue margin tables.

First thing to try

Open Compute Factors and change nothing but the PUE. The hedge ratio moves more than most people expect, which is the entire argument of the denominator work in one slider.

Commodities

Energy complex decomposition

You are long 40 contracts. How many independent bets is that, really?

A position blotter tells you what you hold. It does not tell you what you are exposed to, and in energy those are very different statements — a Henry Hub length and a Waha length are not the same trade, and a PJM-West power position is partly a gas position wearing a heat rate. This tool takes the full listed complex and resolves every contract into dated risk-factor legs, so a book collapses into the handful of factors actually driving it. Gas is regionalized across more than seventy delivery points rather than treated as one curve; power runs ISO to hub to zone with its associated gas hub and heat-rate band attached, which is what makes the spark spread fall out rather than needing to be assembled; crude is split by region, grade and application. Open interest and volume ride alongside, because a factor you cannot get out of is a different risk from one you can. The cross-exchange offset detection is the part most books need: a long on ICE and a short on NYMEX that look like two positions are frequently one, and margin should be netted against that.

What's inside
430+ risk factors470+ contractsICE EU · ICE US (IFED) · NYMEX-CME · Nodal70+ gas delivery pointsPower ISO → hub → zoneCrude by region × grade × applicationRefined products, NGLs, coal, emissionsOpen interest & volumeFive worked portfoliosScenario riskCross-exchange offset detection
Runs on

Exchange contract specifications and settlement records across four venues. The reference engine is checked against a Python implementation, 32 of 32 cases in agreement.

First thing to try

Load one of the five worked portfolios before building your own. They exist to show what a decomposed book looks like when the answer is interesting — the offsets are usually not where you would guess.

Predictive · Perps

Perps × predictive margin analytics

What does it cost to hold a perp and an event contract on the same underlying?

Two instruments, one underlying, and no venue that margins them together. A perpetual future carries linear delta; a binary event contract carries digital delta that is largest exactly where the perp's is flat, near the strike and near expiry. Held together they can be a hedge or a doubling-down, and which one it is depends on the strike, the tenor and the correlation assumption — none of which a per-instrument margin calculation can see. This tool puts both in one portfolio and prices the risk properly: an eight-cluster correlation-aware framework run through a 5,000-path Monte Carlo, with the implied-vol surface solved from the binary strip itself rather than imported. The venue-aware comparison is the practical output — the same book carries materially different margin depending on where each leg sits, and that difference is often larger than the edge being traded.

What's inside
BTC and SPX underlyingsPerp delta vs binary digital deltaPnL curve & net deltaImplied-probability extractionImplied-vol solver + surface8-cluster correlation framework5,000-path Monte CarloVenue-aware margin comparisonLive Kalshi feed
Runs on

Live Kalshi market data with static fallbacks, against published venue margin methodologies.

First thing to try

Put a perp against a binary strip straddling its strike, then push the correlation assumption. Watching portfolio margin move while neither leg changes is the point of the whole framework.

Predictive · Sizing

Portfolio allocator

Given an edge and a margin constraint, what size is defensible?

Sizing and margin are usually solved in the wrong order — pick the position, then discover what it costs to carry, then cut it. This tool solves them together, and it is deliberately the same engine as the margin framework above rather than a second model that would eventually disagree with the first. You supply edge views per cluster; it computes Kelly weights from them, then scales the whole allocation globally so that correlation-aware portfolio margin lands on a bankroll utilization target you set. The ranking that matters is not edge, it is edge per marginal margin dollar — which routinely reorders a book, because the highest-conviction position is often the one that consumes the most capacity to hold.

What's inside
Per-cluster edge viewsKelly and fractional-Kelly weightsGlobal scaling to a utilization targetCorrelation-aware margin constraintRanked by edge per marginal margin dollar
Runs on

The same eight-cluster margin engine as the analytics tool, consumed directly as an input rather than reimplemented.

First thing to try

Enter your real edge views, then halve the utilization target. If the ranking reorders, the book was margin-constrained rather than conviction-constrained — worth knowing before it matters.

Live scanners

Three tools pointed at one trade each.

All three run the same machinery, because all three are asking one question in different clothes: does a set of prices that must be internally consistent actually agree? A mutually exclusive, collectively exhaustive set of outcomes has to sum to one. A milestone ladder has to be monotone. A series price has to reconcile with its own game-by-game implied probabilities. Where those identities break, something is mispriced — and the arithmetic tells you which side without needing a view.

The honest caveat, stated the same way in every one of them: a violation on a market nobody is making is a documentation artefact, not an opportunity. Thin books, wide spreads and access restrictions eat most of what these surface, which is why every ranking is net of fees on both legs and why displayed size matters more than displayed edge. Read them as instrumentation for how well a venue is functioning, and only then as a trade list.

Kalshi · Polymarket · ManifoldStart here

Cross-venue divergence scanner

The same question priced differently on two venues.

Fourteen question links, matched by hand rather than by string similarity, because the failure mode of automated matching is pairing two contracts whose resolution criteria differ in a way that only shows up at settlement. Ranks by executable edge after fees on both legs — buy YES where it is cheap, buy NO where it is rich, hold the pair to resolution — with quarter-Kelly sizing attached.

Open
Tournament structureLive event

World Cup 2026 arbitrage engine

A team's milestone ladder disagreeing with its own outright price.

Milestone markets — make the round of 32, the round of 16, the quarters, the semis, the final — are conditional probabilities in disguise. Multiply the chain back out and you get an implied outright that can be compared against the outright market directly. Sixteen contenders ranked by that dispersion, with an interactive probability tree and a cross-milestone consistency check that flags ladders violating their own monotonicity before any cross-market comparison starts.

Open
Series structure

NBA Finals cross-venue arbitrage

Series-winner prices inconsistent with the game-by-game ladder.

The same decomposition against best-of-seven math: back out per-game implied probabilities from series prices, and check them against what the individual game markets say. Adds a dutching detector and an LP-optimal portfolio across both venues, with a maker/taker fee model and slippage taken from the actual bid-ask rather than assumed away.

Open

Research companions

Tools that live inside a piece.

These do not have their own homes — each is embedded in the research it belongs to, positioned at the point in the argument where you would otherwise have to take a number on faith. Open the piece and the tool is inline. What follows is what each one makes drivable, so you can go straight to the one that matters.

Reading the output

Four conventions that hold across every tool.

Worth two minutes before you take a number out of any of these. The conventions are the same everywhere, and three of the four exist to make the output look worse than the naive version — which is the point.

01

Fees are charged on both legs, always

Every edge figure in every scanner is net of the fee schedule on each venue involved, and slippage is taken from the displayed bid-ask rather than assumed to be zero. Gross spreads look considerably better and mean considerably less — a large fraction of apparent cross-venue arbitrage is a fee schedule.

02

Sizing is quarter-Kelly, and that is deliberate

Full Kelly is optimal only if your probability estimate is exactly right, and on event markets it is not. The quarter fraction is the standard concession to estimation error; it is exposed as a parameter, and raising it is a decision to be made knowingly rather than by default.

03

Measured, modelled and illustrative are labelled apart

Measured figures come from settlement records or published traces. Modelled figures come from a reproducible calculation stated on the page. Illustrative constants exist to make a structure rankable and are marked as such — they are chosen to order designs, not to price trades.

04

Live feeds fall back rather than fail

The scanners pull live venue data where the API allows it and drop to a dated snapshot where it does not. If a figure looks stale it probably is, and the tool says so — a quiet fallback that pretends to be live is worse than no feed at all.

Where these sit

Each tool belongs to a thread.

The tools are easier to place if you read them alongside the argument they were built for. Each section below sequences its research and its tooling together.

On deck: political-event cross-margining, ETH and SOL underlyings for the predictive margin framework, an on-chain data layer across the portfolio, and a listed-options layer for the GSR perpetual.