Crypto up/down research

What I built

What maker value remains from observed BTC price information after order activation, queue access and inventory are included? I studied this question in historical BTC five-minute up/down markets during May–July 2026, connecting four pieces of work:

• Pricing: derive a Brownian-probit description of binary prices, compare feeds and anchors, and estimate constant and feature-dependent pricing scales.
• Measurement: align asynchronous feeds and order books by information availability, identify recording gaps and model when a new order can enter the queue.
• Decisions: express UP/DOWN inventory as paired collateral and unmatched exposure, then rank book-priced actions using joint-fill EV and inventory rules.
• Evaluation: compare policies on common slots, use chronological blocks and scrambled signals, and reconcile apparent gains with execution assumptions.

Contributions and implementation map

Constructing the pricing model

A Brownian-probit model described broad within-slot variation in market mids. This is a result about market pricing structure, not the identification of an individual maker's quote schedule. The market retained better outcome calibration when model and market disagreed.

I separated the underlying diffusion assumption, the fitted market-pricing scale and the induced binary-price dynamics. The public estimation study rechecks seven fixed candidates on 358 archived feature records, with feed, anchor, lag and feature-availability conventions made explicit. A feature using later within-slot prices can describe a fitted scale without being available for an entry decision. Hybrid model versions combined a slowly updated offset with faster price movements and cached state.

Pricing derivation · Estimation and features

When information can reach an order

The event analysis separates exchange price changes, local feed receipt, book response, order activation and queue consumption. Two corrections materially changed apparent opportunities: alternating ticks from exchanges with different price levels created false spikes, and fills before a maker's 50 ms activation delay were unavailable to that order. Missing book updates also required health masks.

Execution notes trace the historical v1/v2 metadata and balance caches into the maker submission route. That route builds and signs on submission; the shared clients' pre-sign helpers are a separate capability. This distinction keeps the weather prepared-order design and the inspected crypto maker implementation accurate to their actual call paths.

Data and execution engineering · Queue and fill mechanics

Inventory risk and the decision rule

The EV decomposition has four outcomes: both legs fill, only the ask leg fills, only the bid leg fills, or neither fills. Joint-fill dependence, conditional post-fill drift and the remaining inventory determine the value of quoting one or both sides. Prices come from the book; model information informs timing and direction. The implementation selects one-step EV actions with rebalancing and inventory rules.

The binary-risk derivation connects the diffusion model to terminal Bernoulli exposure and examines the Avellaneda–Stoikov analogy. An exact CARA valuation diagnostic was added for the public exposition. It is separate from the historical one-step policy; a solved Bellman or deployed A–S optimizer is not claimed.

EV equations and policy · Binary risk and A–S comparison

Selected findings

• Book response: after Binance moves of at least $5 over 200 ms, 87.05% of non-fizzled first responses agreed with the direction. Median delay among correct confirmations was 275 ms. This measures book reaction, not a trader's capture rate.
• Inventory policy: on 368 common simulator-output slots, the inventory-aware policy improved on the baseline by 68.50 cents per slot, paired 90% interval [39.35, 95.17], at a five-share base clip. The comparison remains conditional on queue access, rebalancing quantity and online-timing assumptions.
• Rejected front quoting: the final static candidate had fresh-data EV(0) of −0.9845 cents per eligible decision moment, 90% slot-cluster interval [−1.626, −0.364], across 312 moments in 193 slots. EV(0) sets additional unobserved fill cost to zero; the candidate failed even at that boundary.

A separate six-slot replay/paper discrepancy exposed harness defects. It provides mechanics evidence, without an economic conclusion about a correctly functioning strategy. An earlier A/B comparison was also invalidated by an execution-assumption audit. These findings changed how I interpreted favorable replay results.

Response measurements · Policy comparisons · Rejected candidates and corrections

Read and reproduce

The working draft develops this argument with equations, figures and experiment records. It is available for feedback; formal submission will be considered later. The repository's reproduction guide provides pricing, queue, risk and policy walkthroughs, plus rechecks of released features and simulator outputs. The D12 verifier checks a stored aggregate; the package does not reconstruct unpublished raw recordings.

The study uses its historical point-settlement specification. Its findings are not transferred to subsequent settlement rules. A positive-EV strategy has not been validated in live trading.

Read the working draft · Download PDF · Open the research repository