By James Teschner, Quantitative Analyst Intern, BMLL
The full order book says much more about how far prices are about to move than about which way they will move. The strongest predictor in this study, the fleeting-order ratio, is only observable because orders in BMLL harmonised Level 3 data carry a persistent order identifier. This article explores how market participants can improve short term volatility forecasting with BMLL harmonised high quality historical data.
Short-horizon volatility forecasts are usually built from trailing volatility, and for good reason. At these time horizons, the recent past is hard to beat. In this study, trailing volatility explains 40 to 66 percent of the variation in realised volatility over the next 30 seconds in every test month, across both calm and volatile periods. A detailed view (l3) of orderbook activity does not replace that forecast, but adds a modest improvement to predictions that holds consistently across every month we tested.
Across 395 trading days and 92 million one-second observations, we compared two versions of the same forecasting model. The first used only trailing realised volatility. The second added ten features built from BMLL's order book data. Holding the model fixed, the version with order book features was more accurate in all 16 out-of-sample test months, with a median improvement of +0.04 R². The gain is modest, but it persists across the full test period, including a genuine volatility shock.
The analysis has two parts. Study 1 measures each signal individually to see what information it contains about future price moves. Study 2 asks whether those signals improve a volatility forecast when used together. We also separate the signals by the data required to compute them, from a snapshot of the book to the full message stream with order identity attached.
For our study, we use BMLL Level 3 data harmonised schema that carries detailed features and tracks orders at the message level with a single order identifier.
From raw Level 3 events, we built the forecasting features as part of a 21-signal library (Table 1), calculated on a one-second grid for ten of the most actively traded US stocks and ETFs. The forecasting target is realised volatility over the following 30 seconds.
Level 1 provides the touch bid and ask resting liquidity, Level 2 adds depth at each price level in the order book, giving a snapshot of the order book at a particular moment. Level 3 instead records the sequence of events that produced that book, including every insert, amendment, cancellation and execution. Because BMLL’s order identifier persists throughout the order lifecycle, an individual order can be followed from arrival, through any subsequent amendments, until it is either filled or cancelled. Departures are identified as executions or cancellations, while the market-state field allows continuous trading to be separated from auctions. BMLL harmonises these messages into a common schema across venues while retaining the original exchange message and sequencing. This event-level history is what makes measures such as the fleeting-order ratio possible, since an order that arrives and is cancelled within 100 milliseconds can disappear entirely between snapshots.
Study 1: The order book can help forecast ensuing moves magnitude, not direction
Before combining the signals in a forecasting model, we first measured each one individually to determine whether it contained information about the direction or magnitude of the next price move. For all 21 signals, we computed a rank information coefficient (IC), using Spearman correlation against forward returns at horizons ranging from 5 seconds to 15 minutes. Each signal was tested both against the signed forward return, which captures direction, and against the absolute forward return, which captures the size of the move regardless of direction. An IC of 0.20, for example, means that the ranking of observations by the signal has a 0.20 correlation with the ranking of the subsequent moves by size.
Because adjacent one-second observations overlap heavily, the full row count would overstate the amount of independent evidence in the sample. We therefore compute ICs separately for each instrument and trading day before aggregating the results, so the significance of the findings reflects their consistency across roughly 395 trading days rather than millions of highly overlapping observations.
Figure 1 shows a clear difference between direction and magnitude. No signal at any horizon has a mean IC above +0.029 against the signed forward return, while the strongest signals against the absolute forward return have ICs at least six times larger. The order book therefore contains considerably more information about the size of an imminent price move than about its direction.
Download chart
Figure 1: Mean Spearman IC per (instrument, day) at the 30-second horizon, ten instruments over 395 trading days. Left panel: signals ranked against the signed forward return. Right panel: signals against its absolute value. Note the identical x-axis scale on both panels.
The strongest signal is the fleeting-order ratio at +0.20, which measures the share of recently departed orders that left less than 100 milliseconds after arriving. Because the metric depends on matching an order’s arrival to its departure, it can only be computed from data that preserve order identity across the full lifecycle. Book slope, which measures how quickly depth accumulates away from the touch, carries −0.16, with deeper, flatter books preceding calmer periods, while quoted spread, a Level 1 quantity, carries +0.12. The +0.20 IC of the fleeting-order ratio is notably larger than the other signals in the study and maintains the same sign across all ten instruments.
To test whether this relationship was specific to a particular market regime, we tracked the signals through the full 19-month window shown in Figure 2. Each of the four signals shown maintains its sign in essentially every month, while the fleeting-order ratio remains positive across all ten instruments and shows little evidence of decay over the period. Queue-refresh latency is similarly consistent in the opposite direction, ranging from −0.05 to −0.13, with faster-refilling queues tending to precede larger moves. In this sample, rapid replenishment therefore appears to characterize a more active book rather than necessarily a more resilient one. Like the fleeting-order ratio, queue-refresh latency requires the message stream rather than a snapshot, since the calculation must observe the execution and distinguish a genuine new insert from an amendment.
Download chart
Figure 2: Monthly mean daily IC against the absolute forward return at 30 seconds, matching the horizon used throughout. The pattern holds from January 2025 through July 2026.
Study 2: Order book features added predictive power beyond trailing volatility in every test month
Measuring signals individually establishes what information they contain, but not whether that information improves a forecast. To test this, we ran the same gradient-boosting model with two sets of inputs: trailing realised volatility alone, and trailing volatility plus ten order book features identified in Study 1. Every feature summarises the preceding 30 seconds and the target is realised volatility over the following 30 seconds, making trailing volatility a deliberately strong matched-horizon baseline. The model was trained on a rolling three-month window and tested on the following month, producing 16 out-of-sample test months from April 2025 through July 2026.
Adding the order book features improved the forecast in all 16 test months, with a median increase of +0.04 R² and an increase in mean R² from +0.55 to +0.59, equivalent to a relative improvement of approximately 7%.
We ran two additional comparisons to understand how much of this improvement depended on the baseline or model class. Relative to a linear persistence model using the same trailing volatility measure, the full gradient-boosting model performed better in 13 of the 16 months by a similar median margin, suggesting that the result is not specific to the way persistence is modelled. The three exceptions occurred when volatility moved beyond, or close to the limits of, the range seen in the three-month training window, causing the tree-based forecasts to saturate while the linear persistence model could continue to extrapolate. A linear model using the same ten order book features also improved on its own baseline, producing a median increase of +0.02 R² compared with +0.04 for the gradient-boosting model. This suggests that some of the additional predictive information in the order book is accessible to a linear model, while the non-linear model captures a larger improvement.
Download chart
Figure 3: Walk-forward evaluation, 3-month rolling train, 1-month test, forward 30-second realised volatility. Top panel: out-of-sample R² for both gradient-boosted models and a linear persistence baseline. Bottom panel: the difference between the two gradient-boosted models, which is the lift attributable to the features alone.
Conclusion: Order book features add consistent predictive power beyond trailing volatility
We are not suggesting this constitutes a profitable trading strategy. As we measure returns are mid-to-mid, the analysis is designed to measure predictive information in the order book rather than the profitability of a would be executable strategy. The results show that order book features add a consistent improvement to an already strong short-horizon volatility forecast, with the added features improving performance in every test month. Broader universes, less liquid names and a consolidated cross-venue view are natural extensions of the analysis.
Many of the useful signals in this study can be constructed from standard depth data, while others require the full Level 3 event stream. Within Level 3, persistent order identity makes it possible to follow individual orders through their lifecycle, which is necessary to compute the strongest single predictor in the study, the fleeting-order ratio. Other measures, such as queue replenishment, also rely on the sequence of events captured in the Level 3 message stream rather than on snapshots alone.
The analysis was built entirely in the BMLL Data Lab, including all metrics, models and figures. The underlying pipeline has also been packaged as a reproducible feature-store starter kit, containing the logic for all 21 signals, causality controls and a build engine that can construct the feature store for different universes and date ranges. The study presented here therefore covers only a subset of what can be explored using the underlying Level 3 data.
The starter kit is available to BMLL clients and can be deployed directly within the BMLL Data Lab to run across the universe, resolution and forecasting horizon relevant to their research.
BMLL Level 3 data harmonised schema carries detailed features, and tracks orders at the message level with a single order identifier, enabling market participants to better forecast short term volatility.
How we did it
The analysis uses BMLL Level 3 historical order book data for AAPL, MSFT, NVDA, AMZN, GOOGL, QQQ and TLT on Nasdaq, and SPY, IWM and GLD on NYSE Arca, from 1 January 2025 through 31 July 2026. The sample covers 395 trading days and more than 92 million one-second observations.
We constructed 21 signals from raw Level 3 events on a one-second grid, spanning book shape, order flow and order lifecycle. These include measures of depth slope, symmetry and concentration; multi-level order-flow imbalance, signed trade flow and VPIN; and lifecycle measures such as the fleeting-order ratio, queue-refresh latency, level tenure, queue-age imbalance and an iceberg-activity proxy. Level 3 also allows trade direction to be identified directly from the resting side of each execution rather than inferred from price movements. All features are constructed using only information available at the time of observation, with metrics such as impact and queue refill recorded only once they have been realised.
For Study 1, rank information coefficients are Spearman correlations calculated separately for each instrument and trading day before being aggregated. This avoids treating millions of overlapping one-second observations as independent evidence, with statistical significance instead based on the distribution of daily results. The targets are mid-to-mid forward returns and their absolute values at horizons from five seconds to 15 minutes.
For Study 2, the forecasting target is realised volatility over the next 30 seconds, calculated from the subsequent one-second returns within the same trading session. Each model input summarises the preceding 30 seconds, creating adjacent, non-overlapping feature and target windows and making trailing 30-second realised volatility the matched-horizon baseline. After removing observations without a complete feature set or forward window, the forecasting sample contains 86.4 million observations. Features and targets are standardised by instrument using statistics calculated from the training window only.
The forecasting models are evaluated using a three-month rolling training window followed by a one-month out-of-sample test, producing 16 test months from April 2025 through July 2026. Each fold compares the gradient-boosting model using trailing volatility alone with the same model using the ten selected order book features, while also fitting a linear persistence baseline and a linear model using the full feature set. Model hyperparameters remain fixed throughout the evaluation. The signal shortlist was established during the first quarter of 2026, which falls within the test period, so three of the 16 test months are not fully independent of the feature-selection step.
Data requirement for each signal (Table 1)
Ready to see the market differently?
Discover how BMLL helps trading desks optimise execution and market insight
