Order book microstructure, checksum verified

Book maintenance

The page subscribes to Kraken's book channel at depth 100. The first message is a snapshot of the best 100 price levels on each side. Every later message lists only the levels that changed, each with its new total size, and a size of zero removes the level. Each side is held as a sorted array (bids descending, asks ascending) with binary search for insert, update and delete.

After every update both sides are cut back to 100 levels. The exchange stops sending changes for a level once it falls past the subscribed depth, so a level kept beyond that point would carry a stale size. The checksum covers only the top 10 levels and cannot catch a missed truncation, so the test suite checks it separately against a captured stream.

Checksum

Every book message carries a CRC32 of the exchange's own book. The page builds the same string from its local copy, taking the 10 best asks in ascending price, then the 10 best bids in descending price, each price and size printed with the instrument's precision, decimal point removed, leading zeros stripped. Precision is read from the instrument channel when the connection opens (for BTC/USD, 1 decimal for price and 8 for size).

Formatting is where implementations fail. A size of 0.0157 must print as 0.01570000 and contribute 1570000; printing the shortest decimal form, 0.0157, gives a different string and every checksum fails. On a mismatch the page counts it, unsubscribes and resubscribes the book, ignores updates until the fresh snapshot verifies, and counts the resync.

Top of book

With best bid price Pb and size Qb, and best ask price Pa and size Qa:

The microprice weights each price by the size on the opposite side, so a heavy bid pulls it toward the ask. It is a better guess than the mid at where the next trade will print only to the extent that touch size carries information, which the imbalance chart lets you watch.

Order flow imbalance

Following Cont, Kukanov and Stoikov (2014), every change of the best quotes contributes an event e, where primes mark the previous state:

An unchanged quote contributes its size change. A bid that moves up adds its whole new size, a bid that moves down removes its whole old size, and the ask side mirrors this with the opposite sign. Events are summed over each second of exchange time to give OFI, and the change in mid over the same second is dP. The scatter fits dP = a + b OFI by ordinary least squares over the last 300 seconds:

The paper found a linear relation between order flow imbalance and short horizon price changes for US equities. Here it is tested live on one crypto venue. A second that contains a resync is left out, because its price change has no measured flow behind it. The standard error assumes independent residuals with equal variance; per second crypto returns are heavy tailed and clustered, so read it as a lower bound on the uncertainty.

Realized volatility

The mid is sampled on a fixed grid, taking the last mid at the end of each exchange second, carried forward through quiet seconds. With log returns r = ln(m_i / m_(i-1)) over a window of T seconds:

The window is the last 300 seconds, which gives 300 returns at 1 s sampling and 30 at 10 s. Using 365 days reflects a market that trades every day; the annualized figure rescales five minutes of movement and says nothing about the coming year. If returns were independent and Gaussian, the relative standard error from n returns would be about 1/sqrt(2n), roughly 4% at 300 returns and 13% at 30.

Finer sampling usually reads higher. The observed mid is the underlying price plus noise from tick discreteness and quotes that flicker back and forth, and that noise adds about 2 n η² to the sum of squares for n returns with noise variance η². It grows with the number of returns while the underlying variance does not, so the gap between the two figures is a live estimate of microstructure noise. When the 10 s figure is the higher one, price moved in short runs within the window, which noise cannot produce; with 30 returns that can also be sampling error. The tile compares the gap with twice its standard error before naming either cause.

Limitations

• One venue. Kraken's book is a slice of the market for each pair; other exchanges are not seen.
• Depth 100 only. Nothing past the 100th level on either side is visible, and the heatmap hatches that unobserved range in grey.
• The trade side is the taker side exactly as the exchange reports it.
• Delay includes the offset between your clock and Kraken's. Its level is only as good as your clock sync; its changes over time are more trustworthy.
• OFI, mid changes and realized volatility use exchange timestamps. The heatmap, spread chart, trade flow and trades per minute use arrival time on this computer.
• Everything resets when the symbol changes or the page reloads; nothing is stored. Painting runs at a few frames a second, decoupled from the message rate, and pausing the display stops painting but not ingestion or verification.