Steel: measuring a market that never closes, and refusing to pretend

thingg, LLC · source at github.com/thingg-co/steel · PolyForm Noncommercial · a sibling of Flint, whose paper this one assumes and does not repeat. Steel is a research instrument, not investment advice.

1. A market that never closes

Flint's six principles stand: distributions rather than points, the clock as the only train/test split, one network for the cross-section, defined risk only, everything observable, and one machine that nothing leaves. Crypto adds a seventh, which this paper returns to repeatedly: the simulation pays every cost a real action would, including the costs that equities do not have. A paper book that skipped a withdrawal fee or a ten-minute settlement would flatter the policy on exactly the trades that fail in practice. And an eighth, which gives Section 8 its name: a map that is modeled says so, and says how well it has done.

Five differences from equities drive the rest of the design.

1. No session. There is no close to train through, no overnight gap, no open to wait for. Budgets, limits and refreshes turn over at 00:00 UTC. A gap in the tape is never a holiday; it is an alarm.
2. Many venues, one asset. The same coin trades on twenty exchanges with different quote currencies, fees, pair spellings, and occasionally a different coin under the same ticker. The data layer has to merge, rank and disambiguate before anything is modeled.
3. Derivatives instead of borrowing. There is no stock loan. A short is a perpetual future with isolated margin, which bounds its loss by construction and pays or receives funding every hour or eight. A long can be the same instrument at 1x, which matters because spot retail fees are an order of magnitude higher than perp fees.
4. Money moves on chains. Capital on one venue is not capital on another until a withdrawal clears and a deposit confirms: minutes for a stablecoin on a fast rail, most of an hour for bitcoin on its own chain, with a fee each way.
5. The crowd is visible. Funding, open interest, liquidations, the long/short ratio, the basis and the options dealers' gamma are published by the venues themselves, which gives the policy a direct reading of positioning rather than a proxy.

2. Measuring the market

A trading terminal shows what the venues publish. Steel stores it, so that the model, the dashboard and other programs read the same thing, and so that every number can say where it came from. Six stores, each a module with an API route:

• The tape. Every real print from every enabled venue, active or standby, never the simulator, in a ring per pair: the last prints, volume and delta by eleven size bands, CVD per bar, candles at any timeframe, the footprint (sells at the bid and buys at the ask per price bucket, the point of control, the value area, stacked imbalances, the large prints), and each venue's volume-weighted price against a reference venue, which is the premium.
• The books. Each venue's resting bids and asks from its depth channel, kept within one dollar class at a time (a USDT book merged with a USD one puts the best bid and ask a basis apart), the seven depth bands in dollars per side, the aggregate ladder with venue shares, and a decimated level history for the heatmap.
• The derivatives. Liquidations from Binance, Bybit and OKX (a long liquidated is a forced sell), and open interest, funding, mark and index from those three and Hyperliquid, aggregated with open interest summed in dollars and funding weighted by it. Net positioning reads the open-interest change against the taker side: new longs, new shorts, longs closing, shorts covering; heavy volume on flat open interest adds nothing.
• The options surface. Deribit's BTC and ETH chains every ten minutes: Black–Scholes gamma from each mark's implied vol, dealer gamma by strike, the smile per expiry, the term structure, max pain and the zero-gamma level.
• The calendar and the crowd. The FOMC, CPI and jobs dates as published, the monthly and quarterly expiries computed, anything else from a file; Polymarket's open crypto markets with the crowd's probabilities; Fear & Greed, Reddit, funding, the long/short ratio, open interest and the basis as the crowd overlay reads them.
• The dislocations. Every venue's top of book against every other's, within the same dollar and the same kind, net of both venues' taker fees (Section 7).

Each name is owned by the highest-priority enabled venue that lists it and is fresh; ownership fails over without operator action. Three venues are on out of the box (Coinbase, Kraken and Hyperliquid, the one perpetuals venue); the rest are present but off, because a box in a region a venue blocks, or one that wants two websockets rather than twenty, should start clean. Perpetuals venues rank below every spot venue for price, because a perp is a derivative that funding keeps near spot rather than spot itself. Two rules were learned on the first all-venue run. Quote class. Every venue declares whether its dollar pairs are quoted in USD, USDT or USDC; a standby venue may refresh a name's bid and ask only when its quote matches the active venue's, because a USDT book sat ten basis points above Coinbase's last trade and the policy read that basis as the cost of a fill. Identity. A ticker is not an asset: the same three letters named two different coins on two venues, and the dislocation monitor reported a 300-basis-point arbitrage between them. Steel keeps CoinGecko's top thousand by symbol and the names venues publish, marks a symbol with several ranked coins as ambiguous, and lets such a name pair across venues, or be served at all, only from venues resolved to the same coin. A third rule followed: the simulator never speaks for a name a live venue lists; when every live venue is stale the provider is nobody, so the stall alarm fires instead of a made-up price reaching the labels.

Bars are built only from real trades; a quote or heartbeat updates the last price and the spread but never forms a candle. The taker side comes from each venue's own convention, which differs enough to be worth a table in the source code: Coinbase reports the maker's side, Binance a flag that the buyer was the maker, most venues the taker directly. Stalls are alarms: with no closed market to excuse them, three bar intervals without a candle while prices tick means a trade stream died, and the dashboard and the alerts say so.

3. From measurement to features

Each name contributes 40 features per bar, standardized by a running mean and variance kept per name. Nine come from the bars; thirteen are exogenous signals joined from the signal loops and held between refreshes; four read the resting order book across venues (the imbalance within 0.5%, 2% and 10% of the mid and the depth within 1%); four are forced and confirmed flow (longs and shorts liquidated in the bar, new longs and new shorts the open-interest change confirms); two are the options surface for BTC and ETH (dealer gamma as a share of the chain and the at-the-money implied vol); one is the clock to the next scheduled high-impact event; seven are functions of the bar's timestamp, which matters in a market where the hour of day, the funding settlement clock and the CME session all shape the tape.

The net

SteelNet is FlintNet with a horizon axis. The same trunk, dilated causal convolutions over the 128-bar window per name with a per-name embedding, cross-asset attention across names, and a regime gate over expert heads, feeds H quantile sets and tail pairs instead of one. The first horizon, twelve bars, is the policy's; the extras, 48 and 288 bars, are forecasts the dashboard shows beside it and the policy consults as a veto. The loss weighs each extra horizon at half the policy's:

with w₁ = 1 and w_h = ½ for the extras. Why one trunk and not three models: the extra horizons are there to say whether the hour's call is a scalp into a trend, and a shared representation makes that comparison coherent rather than a disagreement between strangers. Only the first horizon is ever traded. Sizing to the machine is Flint's ladder, with one addition forced by the reference box: Steel shares its GPU with Flint, which holds about fifty gigabytes on weekdays, so the ladder runs under a memory floor and the start refuses, naming the other tenant, when the chosen rung would not fit beside it.

The replay and the universe

A market with no close produces 288 bars a day, every day, and the deep backfill reaches back months, so Steel stores one row per bar with its label, mask, timestamp and links to the bars before and after it, and gathers windows by index at batch time; a window is valid only over time-contiguous bars. Consecutive windows share all but one bar, so this holds sixty-four times the history Flint's window ring did for the same bytes: a year of four hundred names fits in under four gigabytes. The closes the store keeps also give the four-hour and one-day labels for free. Every twentieth bar by timestamp is held out of training for good as the one honest overfitting signal.

There are far fewer liquid crypto pairs than liquid stocks, so there is no reason to curate a small set. In the default liquid mode the universe is a sixty-name core, every holding in the operator's accounts, and every dollar pair the enabled spot venues list whose 24-hour dollar volume clears a floor (one million dollars a day, about 120 Coinbase names today), under a hard cap of 512 that is a limit rather than a target. Stablecoins and the fiat pairs Coinbase lists as dollar products are never modeled. Flint's grace and tenure rules apply, so a thin day never churns the set. The modeling floor is separate from the trading floor, which is twice as high.

4. Learning around the clock

Flint learns in four regimes separated by the session. Steel has no session, so the regimes are separated by budget and by who else wants the accelerator.

• Warmup. A start pages thirty days of candles for the universe from the spot venues, perpetuals venues and the simulator filling only names left empty, and trains one pass over the labeled bars with the held-out check every forty steps, stopping when the held-out pinball stops improving or at a thirty-minute cap. Only then does it forecast.
• Online. Each bar whose forecasts have resolved yields a few steps on recency-biased batches, exactly as in Flint.
• The background trainer. Between bars, up to train_epochs passes over the replay per UTC day, in four-step chunks so a forecast waits one step at most, under the same held-out guard. The budget is bounded by data rather than hours, and resets at midnight UTC.
• The deep backfill. Always on: five-day chunks of history every 150 seconds, each name from the first enabled venue that pages by date range and lists it, with eight training steps per chunk off the event loop.
• The GPU handoff. When Flint is running, or will start within fifteen minutes of its timer, Steel is in shared mode: its base budget, its memory floor, the weekly backtest deferred, and in the lead-up the background trainer idle and the CUDA cache released. When Flint is off, solo mode triples the training budget and drops the floor. The model is not swapped between modes: continuity is the point of online learning.

5. Trust

Calibration is Flint's: an online conformal scale on the predicted band widths that starts wide and earns its way down as the 10–90 band's coverage approaches 80%, and a temperature on the direction logits that starts sceptical. Trust is where Steel differs, because a restart on a market that never closes should not mean hours without a book. The held-out slice of the replay over the last seven days is scored the way live labels are, by direction hit on the calls the policy would have made, and the model is trusted when enough such calls hit; live labels take over as they mature. A scorecard that fails never revokes a trust the live record has earned. The per-name skill gate reads the same held-out calls on that name until live calls fill its record, then ages them out. The net's weight per feature is reported as what the net can read from a column, not what the column causes; a near-zero share says the column is ignored, which is the honest reading of a feature that was added yesterday.

6. Every cost is real

The policy is Flint's composition of auditable gates, each writing its reason into the name's card: the expected move against its uncertainty, the direction's margin over a coin flip, the edge net of every cost, the per-name record, the live spread, the daily dollar volume, the order's own impact by the square-root law and Amihud's illiquidity, the trend and the direction agreeing, and the same side persisting over several bars. Three gates are new: horizon agreement (a one-hour call whose four-hour median points the other way is held), volatility targeting (a position is sized to a daily volatility target over the name's realized daily volatility, floored at a quarter) and cluster caps (no sector carries more than 40% gross, and the book's net bitcoin beta stays within 1.0). The crowd overlay fades positioning that is visible here in a way it is not in equities; a thin edge into extreme crowding is flipped, and the overlay says so.

The paper book

No position may carry unlimited risk. A bearish view is an isolated-margin perpetual short at 2x: the margin posted, notional over leverage and capped as a share of equity, is the most it can lose, because the venue liquidates it at entry × (1 + (1 − maintenance) / leverage) and the book settles that at exactly the margin. A bullish view is, by default, a 1x perpetual long on the perp venue, whose worst case is the stake, the same as spot. There is no naked borrow and no leverage on the book as a whole. Funding accrues on every open perp at the venue's current rate, pro rata per bar: a positive rate pays the shorts and charges the longs.

Every cost a real fill pays

Fees are the venue's, not a flat number. A profile per venue holds its taker and maker rates for spot and perpetuals, its funding interval and its maintenance margin; two settings say which venue each side would trade on; and once a day each configured venue's reader fetches the account's actual fee tier and the book charges that instead. The policy charges two taker legs of the venue a side trades on, the funding a perp would pay while the rate runs against it (never a credit for funding it would earn), the spread, the impact and the Amihud charge against the median before it trades; the book charges the same on its fills. The consequence is deliberate: a retail spot tier never looks tradable at a one-hour horizon, so longs go to the perp venue.

Money between venues

The book keeps its cash on the spot venue and its collateral on the perp venue, split at the start. When a position needs more than its venue holds, the book sends the shortfall, as USDC on the cheapest rail in a table of withdrawal fees and settlement times per asset, charges the fee, and holds the money in flight for the rail's wait. The position does not open until the money lands, and only if the signal is still there when it does. The same table prices what the arbitrage executor refuses: a leg that would need a balance moved first is declined with the transfer's cost in basis points of the trade and its minutes, because a dislocation measured in seconds does not survive a ten-minute deposit. The chart draws every fill where it happened with its side and notional, the live mirror's slippage beside it, and each transfer as a band on the time axis from sent to landed.

7. Cross-venue structure

With twenty venues quoting the same names, the manager's book hook delivers every venue's top of book, active or standby, to a monitor that opens an episode when the best bid across venues stands above the best ask across venues net of both venues' taker fees and a latency allowance, within the same dollar and the same kind, and closes it when the gap is gone. The episode's window, its peak and mean net edge and the top-of-book sizes go to a log and to the dashboard. The monitor measures; it never executes. Cross-stablecoin and spot-against-perp gaps are deliberately not episodes: those are a basis to carry, not a round trip.

An executor exists, off unless switched on and execution is enabled, that sends both legs at once when an episode's net edge clears a minimum, both venues are armed, both legs are the same kind, and the readers' last balances can fund them, flattening a leg whose partner failed. Before anything like that runs with money, the question is what the rules would have captured, and the log answers it. A replay applies the executor's rules by latency rung, from 50 ms to 3 s: an episode counts only when it outlasted the round trip, at its mean net edge rather than its peak, at the smaller of the book's size and the order cap, after a cooldown per name; and because each capture moves base to the buy venue and cash to the sell venue, every two orders' worth of one-way flow on a pair is charged an on-chain rebalance of two stablecoin transfers.

Three things in that table are the finding. First, the over-wide episodes were a quote problem, not an opportunity, and the monitor now refuses them. Second, the median episode lasted 171 milliseconds; a human, or a process on the other side of the internet from the venues, sees under half of them. Third, the gross edge is real but small, about twenty basis points on a $200 order, and the inventory it leaves behind costs more to move back on-chain than the captures earn. At this size the honest number is zero. Larger size raises the edge faster than the rebalancing cost, and inventory pre-positioned on both venues removes the transfers, which is what an arbitrage desk is.

8. Honest maps

The order-flow terminals draw two heatmaps that traders act on: where liquidity rests, and where leverage would be force-closed. The first is measured and the second is not, and the honest thing is to say so on the map itself. Steel's book heatmap is the mean resting dollars per price row per column from the level snapshots the venues' depth feeds showed; its label says what rested, not what would fill. Steel's liquidation map is a model: when open interest rises in a bar, the new notional is split by the taker side and projected to each leverage band's liquidation price from the bar's close; when open interest falls the clusters shrink in proportion; when price trades through a level the cluster is spent. No venue publishes aggregate liquidation levels, so the only check the map has is against the forced orders the venues do publish, which is what the calibration score reports, with its tolerance in basis points, beside the map wherever it appears. The options surface is measured from Deribit's marks and modeled through Black–Scholes; the dealer-gamma figure is labelled per 1% move, and max pain and zero gamma are what the chain implies, not where price will go.

The same rule applies to the model's own output. The forecast fan is a calibrated distribution with its coverage tracked; the calls drawn on the chart are filled when they hit and hollow when they missed; the paper book's fills and the live mirrors are drawn where they happened, with the slippage; a feature's weight is called what it is. A dashboard that only showed the hits would be a terminal, not an instrument.

9. From paper to live

Orders exist in one package, behind one gate, and every switch defaults to off. Execution must be enabled; a venue trades only when armed with a trade key of its own, separate from the read-only key; every cap (per order, per name, per venue, per day, exposure, leverage) must be set, because an unset cap is zero; the box clock must be within a second of the venues' server time; the kill switch, a flag, a button or a file, refuses everything; a reject starts a cooldown; and every request and answer goes to an audit log. Clients default to a dry run that signs and logs without sending. Balance readers are a separate package that can only read the endpoints each declares, and a test greps every path-like string in both packages for order, withdrawal and transfer words. The ticket is the operator's: editable, with its limit, stop and target as draggable rays on the chart, and a choice of sending it as one order or as a TWAP, a scale ladder or a chase; every child goes through the same gate.

Promotion is a checklist the operator reads, never a rule the code applies: the model trusted, the weekend backtest scored over enough days with a positive net and an adequate hit rate, the paper book running for enough days with an adequate Sharpe, drawdown and return. Once live, every paper fill becomes an order of a fraction of its notional through the client of the venue that side trades on, the book keeps what it expects to hold, reconciles it against the readers' positions on every poll, reports slippage against the paper price, and demotes itself at once on the kill switch, a drawdown past a limit or a drift past a tolerance. A stop cancels every open order; a panic button closes every position with reduce-only orders; a live book not heard from for half an hour comes back demoted. Alerts carry fills, calls, refusals, stalls, demotions and the kill switch to a phone, and the operator's own series alerts ride beside them.

10. Evaluation

Live scoring is Flint's: every forecast is scored an hour after it is made on direction, band coverage and pinball loss, per name and in aggregate, and the paper book's record is the primary evidence. The held-out scorecard is the second, recomputed on demand over the last seven days and written to a daily scorecard the Model view shows over time beside calibration by name, the drift of each standardized feature over the last day, and the net's weight on it. The walk-forward backtest is Flint's too, with days that are UTC days and a Sharpe that annualizes over 365, run weekly in the thinnest tape, Saturday 00:00 UTC, when it owns the GPU and Flint is off; the policy sweep over its stored forecasts proposes gates without applying them. Whether the forecasts carry economic edge after the costs of Section 6 is the open question the backtest exists to answer, and this paper will carry the number when there is one.

11. Limitations

• The forecasts' economic edge after costs is unproven. On a CPU development box the net is too small to say anything; on the reference GPU the weekly backtest and the paper record are the evidence, and both are young.
• The liquidation map is modeled, and its calibration depends on the three venues that publish forced orders; Hyperliquid publishes none, so its share of the leverage is never scored.
• The transfer table is typical values, not a feed. Venues change withdrawal fees and processing times; the override file is the remedy until a reader fetches them.
• The paper book fills at the touch with the spread charged but without the square-root impact the policy gates on. At the book's size and the names it trades the difference is small; it is the one cost the fill does not pay.
• Identity resolution covers the top thousand coins and the names Coinbase publishes. A collision below that rank, or on a venue that publishes no names, can still pair two different coins; the impossible-gap rule catches the gross cases only.
• The dislocation log was gathered from one site. Latency to each venue, and therefore the share of episodes a real executor would see, depends on where the box sits.
• Every reader and order client was exercised live with dummy credentials up to the venue's "invalid key" answer, which proves the request shape, not the response parsing, and nothing has been sent to a live matching engine.
• The event calendar is the published 2026 dates and the computed expiries; token unlocks and listings are not in it.

12. Future work

1. Pairs as a conversion graph. A pair is an edge between two assets on a venue; a rail is an edge between venues; a router prices the path, and the book holds assets rather than symbols. Most of the transfer and quote-class machinery in Sections 2 and 6 is this graph seen one edge at a time.
2. Real exogenous history in the replay, so the thirteen signal columns are learned from months of data rather than from the live bars since the start.
3. Impact on the paper fill, so the book pays what the policy already charges.
4. Withdrawal fees and times from the venues' own endpoints, replacing the static rails.
5. Funding as a target. The funding rate is forecastable and paid every hour or eight; a carry book that holds the basis with bounded risk is a second strategy the same data supports³.
6. Meta-labeling over the outcome records the book already keeps, sizing by the probability that acting on a call pays after costs⁴.
7. The tape's delta, the premium and the basis as feature columns, each a deliberate reset of the net, with a permutation test for what each contributes.

References

1. I. Makarov and A. Schoar, "Trading and Arbitrage in Cryptocurrency Markets," Journal of Financial Economics 135(2), 2020. Article
2. I. Gibbs and E. Candès, "Adaptive Conformal Inference Under Distribution Shift," NeurIPS, 2021. arXiv
3. A. Schmeling, M. Schrimpf and K. Todorov, "Crypto Carry," BIS Working Paper 1087, 2023. Paper
4. M. López de Prado, Advances in Financial Machine Learning, Wiley, 2018: triple-barrier labeling and meta-labeling.
5. Y. Amihud, "Illiquidity and Stock Returns: Cross-section and Time-series Effects," Journal of Financial Markets 5(1), 2002.
6. R. Almgren, C. Thum, E. Hauptmann and H. Li, "Direct Estimation of Equity Market Impact," Risk, 2005: the square-root law.
7. R. Koenker and G. Bassett, "Regression Quantiles," Econometrica 46(1), 1978: the pinball loss.
8. Flint: design of a continuously learning intraday market model, thingg, 2026. Whitepaper

Appendix: principal settings and the data API

Revision history: v0.1, October 11, 2026, first publication, organised as Flint's paper. v0.2, October 13, 2026, reorganised around measurement and cost honesty; the order-flow stores, the honest-maps rule, the 40-column feature set, the data API and the gated algorithms added; every section opens with a measurement and its date; the appendix is generated from the code by steel appendix.