Glossary¶

Brief definitions of the market-microstructure jargon used throughout ob-analytics. Each entry links to the relevant API or concept. For a from-scratch introduction, start with the tutorial chapter From a price to an order book.

Exchange mechanics¶

Exchange / bourse : A venue where strangers trade a standardized instrument by posting firm, standing offers instead of haggling pairwise. The public list of those offers is the order book.

Matching engine : The exchange's neutral component that pairs compatible buy and sell orders under fixed, published rules and reports the resulting trades.

Continuous double auction : The market design run by modern exchanges: both sides post offers ("double"), and matching happens the moment offers become compatible, all session long ("continuous") — rather than at a scheduled auction time.

Price–time priority : The standard matching rule: better-priced orders trade first, and at the same price, earlier arrivals trade first. The within-price queue it creates is what Level 3 data (and the queue engine) lets you reconstruct.

Market data levels¶

Level 1 (L1) : The top-of-book summary: best bid, best ask, and last trade. What brokerage apps and tickers show as "the price".

Level 2 (L2) — market by price : The full ladder of price levels with aggregate size at each level. Individual orders are summed away.

Level 3 (L3) — market by order : Every individual order with its own identity and queue position — enough to replay arrivals, cancellations, and fills exactly. Also called market-by-order (MBO) data. This is the resolution ob-analytics reconstructs from Bitstamp and LOBSTER feeds.

Best bid / best ask : The highest standing buy price and lowest standing sell price. The ask minus the bid is the spread; their average is the mid-price.

Last trade : The most recent execution's price — the number headlines call "the price", which can move without any trade (see the flash example in the tutorial).

Order book mechanics¶

Limit order : An instruction to buy (bid) or sell (ask) at a specified price or better. Sits in the book until matched, modified, or cancelled.

Market order : An instruction to execute immediately against the best available counter-side liquidity. Modeled in this package as a limit order whose price crosses the spread on arrival.

Maker / taker : The maker is the resting side of a trade (the limit order that was already in the book); the taker is the aggressive side that crossed the spread to consume it. See the maker_event_id / taker_event_id columns documented in Data Contracts.

Spread : Best ask price minus best bid price, in ticks — see Tick size below. Extracted from the depth summary via get_spread.

Mid-price : (best_bid + best_ask) / 2, in ticks — see Tick size below. Reference price for measuring order aggressiveness in basis points.

Tick size : The instrument's minimum price increment, in quote currency. Every price column is stored as a whole number of ticks (int64), not a float in the quote currency; multiply by PipelineConfig.tick_size to recover the quote-currency price.

Basis point (BPS) : 1/100 of a percent. The depth summary bins liquidity into rings of depth_bps width around the mid-price; see PipelineConfig.depth_bps.

Order classifications¶

Produced by set_order_types, which assigns one of six categories.

Unknown : The initial, unclassified state. Remains on any order that fits none of the other categories once classification finishes; set_order_types logs a warning when this happens, since it signals a classification gap rather than a normal outcome.

Pre-existing : An order first seen part-way through the stream, with no created row — the opening book at capture start, or a hidden execution. Structurally unclassifiable rather than a classification failure, so it gets its own category instead of falling back to unknown.

Resting limit : A passive limit order that sits in the book and is eventually filled or cancelled without ever crossing the spread.

Market : An order that crosses the spread on arrival and executes immediately.

Market-limit : A limit order that crosses on arrival but, after partial fills, comes to rest as a passive order at a price inside the book.

Flashed-limit : A limit order that is created and cancelled within a very short window without ever filling. Common in HFT quote-stuffing patterns.

Hidden liquidity¶

Implemented in hidden_liquidity.

Hidden order : An order the venue will match but does not show in the visible book. LOBSTER marks an execution against one as event type 5, with order id 0.

Iceberg order : An order that shows a small displayed peak and keeps the rest in reserve. Each visible part is a slice. When a slice is filled out, the venue shows the next one as a new order at the same price and side.

Refill : The new slice that appears after an iceberg's slice is filled out. detect_icebergs looks for refills within max_delay (one millisecond by default) and chains them into one suspected iceberg.

Trade against a hidden order : A trade that printed strictly inside the visible spread, so no visible order rested at its price. Found by hidden_trades.

Flow toxicity¶

Implemented in flow_toxicity.

VPIN — Volume-Synchronized Probability of Informed Trading : Easley, López de Prado, & O'Hara (2012). Bucket trades by equal volume, classify each bucket as buy- or sell-driven, and report the rolling absolute imbalance. High values (≳0.7) signal informed-trader pressure. The usual bucket size is average daily volume ÷ 50 (vpin_bucket_volume).

Kyle's lambda (λ) : Kyle (1985). Slope of Δprice ~ signed_volume regression over a rolling window. Higher λ → less liquid market (more adverse-selection cost per unit of order flow). Returned as a KyleLambdaResult with the regression DataFrame, a bootstrap confidence interval, and a significant flag that is False on too few windows or |t| < 2.

Order flow imbalance (OFI) : Per-window net buy-minus-sell volume normalised by total traded volume. A short-horizon proxy for directional pressure.

Bars¶

Implemented in bars.

Bar : A summary of a run of consecutive trades: open, high, low, close, volume, and the microstructure columns that go with them. The boundaries come from a bar rule.

Clock bar (OHLCV) : A bar covering a fixed span of the clock — the familiar 1-minute or daily candle.

Tick bar : A bar covering a fixed number of trades. "Tick" here is the market-data sense of one printed trade, not the price increment.

Volume bar / dollar bar : A bar covering a fixed amount of traded size, or of price × size. Sampling by activity rather than by the clock puts the same amount of market in each bar, which brings bar returns much closer to being independent and identically distributed.

Imbalance bar : A bar that ends when signed size drifts a set amount away from where the bar opened, in either direction (López de Prado). A burst of one-sided flow closes a bar; balanced trading stays inside one.

VWAP : Volume-weighted average price — a bar's turnover divided by its volume. The average price actually paid, rather than the average of the prices printed.

Turnover : Price × size summed over a bar: the value that changed hands, in the quote currency. Also called notional or dollar volume.

Signed volume : Buyer-initiated volume minus seller-initiated volume. The net direction of the flow in a bar, and what an imbalance bar accumulates.

Transaction cost¶

Implemented in cost.

Effective spread : What a taker actually paid to cross, measured from the mid-price the trade crossed: 2 * D * (price - mid), where D is +1 for a buy and -1 for a sell. Doubled so it compares with a quoted spread, which also spans both sides of the mid. Unlike the quoted spread it is a property of trades, not of the book, so it reflects where in the book people actually traded.

Realized spread : The part of the effective spread the liquidity provider kept, measured one horizon after the trade: 2 * D * (price - mid_later). A negative realized spread means the provider lost money on the trade — the flow was informed.

Price impact : The rest of the effective spread: how far the trade moved the market, 2 * D * (mid_later - mid). Effective spread = realized spread + price impact, exactly, trade by trade.

Horizon : The wait between a trade and the mid-price the realized spread is read against. Five minutes is the equity convention; a fast tape needs less, because unrelated price moves are charged to the trade as impact. There is no neutral value, so the horizon is reported with the number.

Amihud illiquidity : Amihud (2002). The price move a unit of turnover buys, |return| / turnover over a window. High means thin: a small amount of trading swings the price. Needs no quotes and no aggressor side.

Roll's implied spread : Roll (1984). The spread implied by bid-ask bounce, 2 * sqrt(-cov) over the lag-1 autocovariance of trade price changes. It assumes the bounce is the only thing moving the price, which fixes the lag-1 autocorrelation of those changes at -0.5 — the diagnostic returned beside the estimate. On a tape that is sparse relative to how fast the instrument moves, the price change between trades is mostly efficient-price movement rather than bounce; the autocovariance then comes out positive, the estimate does not exist, and it is reported as NaN rather than hidden. When it lands negative by chance the estimate exists but means nothing, which is why the autocorrelation matters more than the root.

Feature table¶

Implemented in features.

Feature table : One tidy table for a model or a study: a point in time on each row, a microstructure feature in each column. Every row is stated as of one instant, and everything in it was known at that instant.

Feature : One measured column set of that table. Registered under a name, so a measurement of your own becomes a column with no edit to the package.

Look-ahead : Using data from after a row's instant to compute that row. It flatters a model in testing and cannot be repeated in trading, which is the most common way a backtest comes out wrong. The feature table has none by construction: the trade columns cover the bar only, and the book columns are a backward as-of join. : The test for it is truncation. Cut the inputs short and recompute; a table free of look-ahead reproduces the rows that survive exactly, because none of them ever read the data that was removed.

As-of join : Matching each row to the last observation at or before its instant, rather than to one sharing its key. How the book state reaches a row whose instant falls between two book snapshots.

Target : What a model predicts, usually the return over the bar after the row. It looks forward, so it is not a feature and the table does not hold one: build one with a negative shift.

Trade imbalance : The signed share of a bar's volume, signed_volume / volume, from -1 (every trade a sell) to +1 (every trade a buy).

Realized volatility : The standard deviation of returns over a trailing window. Per bar rather than annualized: only clock bars span equal amounts of time, so there is no one factor that would scale it to a year.

Close location : Where a bar's close fell within its own high-low range. Not a built-in feature; the worked example of writing one.

Pipeline concepts¶

Canonical events / trades frames : The venue-independent DataFrames every pipeline stage consumes, defined by the column contracts in Data Contracts. Once frames pass the validators, downstream code cannot tell which venue they came from.

Loader : Produces the canonical frame for a source, via create_loader. An L3 source's create_loader returns an EventLoader, whose load() returns a validator-passing events frame; an L2 source returns a DepthSource, whose load() returns the depth frame directly, since a price-level feed has no per-order events to fold. One per venue dialect.

Trade source : The companion protocol for executions: load(events, source) returns the canonical trades frame with maker/taker attribution.

Metric : A measurement taken from a finished run — registered under a name in METRICS, computed on demand by result.metric(name), and drawn as a level-less plot under that same name. See Metrics and Extending.

Data formats¶

Matched book : An L3 feed produced by the venue's own matching engine (LOBSTER, exchange MBO feeds): bids can never rest above asks, so an uncrossed book is a guaranteed invariant of the data.

Diff feed : An L3 feed reconstructed from a public placement/cancellation stream (the Bitstamp feed used here). It can contain genuinely crossed resting orders; order_book() replays such feeds faithfully — a crossed book in the output is a property of the feed, not a reconstruction bug. See the pitfall note in L1 → L2 → L3 and the Data quality explainer, which shows how to measure the crossing (audit) and uncross it for display (uncross=).

Bitstamp CSV : One row per order event with columns id, timestamp, exchange_timestamp, price, volume, action, direction. The pipeline also expects a sibling trades.csv with live-trade columns (trade_id, timestamp, price, amount, buy_order_id, sell_order_id, side, …) — see scripts/collect_bitstamp_btcusd.py.

LOBSTER : Lim-Order-Book-System-The-Efficient-Reconstructor data set (lobsterdata.com). Provides paired message and orderbook CSV files; integer prices in ten-thousandths of a dollar; timestamps in seconds-after-midnight. Event types 1–7 cover submissions, cancellations, executions of visible/hidden liquidity, cross trades, and trading halts.