A technical field guide to systematic trading. Compare objectives, equations, data dependencies, execution architecture, and the failure modes that separate a research prototype from a trading system. Every entry is classified in six letters and carries its evaluation protocol; 12 of the 16 carry the published evidence against them, and 15 of the 16 run a small seeded computation in the page.
16 STRATEGY FAMILIESSIX-LETTER CLASSIFICATIONPRIMARY-SOURCE REFERENCESSYNTHETIC LABS, COMPUTED IN THE PAGE
Every entry carries one letter per position, in a fixed order: role, model class, learning task, decision horizon, data tier, evidence grade. TUNDBE is multi-horizon momentum: trend & momentum, rules & filters, none, daily to weekly, bars, established. The last letter grades the public literature on the method, not any result of the studio or its clients.
The code is a Mr. Piron convention. Its positional layout is borrowed from ISO 10962 (CFI), which classifies financial instruments in six letters. It is not an industry standard.
Filter by any position of the code. Filters combine.
1Role
2Model class
3Learning task
4Decision horizon
5Data tier
6Evidence grade
SYSTEMATIC TRENDHours to months
/01
Multi-horizon momentum
Combine independent trend horizons, size positions by estimated volatility, and constrain correlated exposure. A transparent research baseline for liquid directional markets.
Fixed lookback rules on bars; nothing is fitted beyond the lookbacks and a volatility floor. Established: the cited studies document the effect in every one of 58 futures and forward contracts over more than 25 years and, from an independent group, across four asset classes with series reaching back to 1800.
Evaluate net portfolio utility across forward periods, subject to gross exposure, volatility and turnover limits; the lookbacks are research parameters, not universal constants.
SYMBOLS
Symbol
Meaning
Unit or domain
ri,t−k
Return of instrument i, k bars before t
fraction per bar
Lℓ
Length of lookback ℓ
bars
aℓ
Weight of lookback ℓ in the combined signal
dimensionless research parameter
σ^i,t
Volatility estimate available at t
fraction per bar
si,t
Combined trend signal; each horizon is squashed by tanh before weighting
dimensionless, bounded
σfloor
Lower bound on the volatility used for sizing
fraction per bar
wi,t
Position weight before portfolio constraints
proportional; scaled to the risk budget
SYNTHETIC · SEED 20261106
What costs, a no-trade band and a fill delay leave of a planted trend
Generator120 seeded paths of 756 bars. The drift follows a three-state chain (+0.12σ, 0, −0.12σ per bar, persistence 0.98) under unit Gaussian noise. The signal is the card formula with lookbacks 8, 32, 96 and an exponentially weighted volatility (decay 0.94). Costs are quoted for σ = 100 bps per bar.
Computed in your browser when the page loads.
01Signal inputs
Lagged total or excess returns
Rolling volatility and covariance
Contract roll and financing estimates
02Data specification
Point-in-time instruments and corporate actions
Separate signal-adjusted series from tradable futures prices
03Execution & infrastructure
Participation caps and spread-aware sizing
Contract, lot and exchange-calendar handling
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITForward periods only
Choose lookbacks on an early span and evaluate on later, untouched periods. The lookbacks are research parameters, not constants.
2
LEAKAGETradable series
Compute signals on adjusted series and fills on the contracts that actually traded, with roll and financing costs.
3
BASELINEBaselines to beat
A single-horizon signal and an equal-risk long-only book on the same instruments.
4
STRESSReversals and costs
Sharp reversals, doubled costs and a lagged volatility estimate.
5
REFUTATIONWhat refutes it
Net of costs, the multi-horizon book does not improve on the single-horizon baseline in forward periods, or the result rests on a few instruments.
04Known failure modes
Whipsaw in range-bound markets
Hidden concentrated macro bets
Unrealistic roll fills or leverage after calm periods
COUNTER-EVIDENCEWhat the literature says against it
Each finding is worded as the cited source reports it and links to that source.
LIMITATION
The reading: A trend persists for as long as it is held.
Moskowitz, Ooi and Pedersen find that the effect persists for about a year and then partially reverses over longer horizons.
Model a tradable residual between related instruments. Test the economic relationship and residual dynamics before implementing entry, exit and hedge rules.
A cointegration regression and a z-score rule, with the hedge ratio optionally carried by a Kalman filter. Active: Gatev, Goetzmann and Rouwenhorst document the returns of a pairs rule on daily US equity data in a peer-reviewed study; whether they survive a given market, period and cost estimate is open, and the entry lists broken linkages and spurious pairs among its failure modes.
Trade a stationary residual with an explicit inventory horizon; optimize entry/exit thresholds only within training folds after accounting for both legs.
01Signal inputs
Aligned tradable prices
Rolling or state-space hedge coefficients
Spread residual and lag-one persistence
02Data specification
Synchronized quotes and action-adjusted history
Short availability, funding and borrow history
03Execution & infrastructure
Two-leg fill and hedge management
Unhedged exposure limits and timeout rules
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITTwo separate holdouts
Select pairs on one span, select thresholds on a second, and evaluate on a third.
2
LEAKAGETraining-window estimates
Intercept, hedge coefficient, residual moments and persistence are estimated on training data only.
3
BASELINEBaselines to beat
A static in-sample hedge ratio and a pair chosen by correlation alone.
4
STRESSBreaks and both legs
Structural-break checks, borrow recalls and the cost of both legs, including legging risk.
5
REFUTATIONWhat refutes it
The residual stops reverting out of sample, or the selected pairs are no better than the number of pairs searched would produce by chance.
Rank instruments using interpretable factors, boosted trees or other supervised models. Translate scores into a constrained portfolio instead of equating prediction accuracy with profit.
Gradient-boosted ranking on point-in-time panels, rebalanced daily to monthly. Active: Gu, Kelly and Xiu compare machine-learning return forecasts on US equities in a peer-reviewed study and find trees and neural networks ahead of linear methods; results in this family depend on universe, period, and costs.
Learn pairwise relative-return ordering; construct exposure-neutral weights with explicit sector, position, liquidity and turnover constraints.
01Signal inputs
Point-in-time price and fundamental factors
Forward excess-return labels
Sector, beta and liquidity exposures
02Data specification
Historical universe membership including delistings
Publication-time and revision-aware fundamentals
03Execution & infrastructure
Rebalance netting and short eligibility
Trade sizes capped by available liquidity
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITDate-grouped rolling folds
All instruments of a date stay in the same fold; folds move forward in time.
2
LEAKAGEPurge the label interval
Remove training rows whose forward label overlaps the validation period; use fundamentals as first published.
3
BASELINEBaselines to beat
A linear combination of the same factors and an equal-weight universe including delisted names.
4
METRICRank and portfolio metrics
Rank IC, top-minus-bottom spread and net turnover, after sector and beta neutralization.
5
REFUTATIONWhat refutes it
Rank IC does not survive the purged folds, or the spread disappears once turnover costs and liquidity caps are applied.
Infer latent volatility or liquidity states and condition exposure on filtered probabilities. Adapt risk budgets or model selection using observations available at the decision time.
A hidden-state filter whose output is a state probability, used to scale exposure. Active: Hamilton’s 1989 Econometrica paper introduced the Markov-switching filter that the cited statsmodels documentation implements, while its value as a trading controller depends on how few transitions the sample holds.
Fit state and transition parameters by likelihood; evaluate whether the resulting controller improves a prespecified forward portfolio objective.
01Signal inputs
Lagged returns and realized volatility
Liquidity features and rolling correlations
State transition probabilities
02Data specification
Causal feature construction
Enough history to observe transitions and stress periods
03Execution & infrastructure
Hysteresis to reduce rapid switching
Risk caps applied outside the probabilistic controller
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITExpanding re-estimation
Fit states and transitions on data through t and apply them after t.
2
LEAKAGEFiltered, never smoothed
Signals use filtered probabilities. Smoothed states use later observations and are kept for retrospective analysis.
3
BASELINEBaseline to beat
A simple realized-volatility threshold with the same hysteresis.
4
STRESSRelabeling and churn
Check that state labels keep their meaning after each refit and count controller switches.
5
REFUTATIONWhat refutes it
The controller does not improve the prespecified forward objective over the volatility threshold, or the sample holds too few transitions to estimate them.
04Known failure modes
Smoothed posteriors used as a signal: hmmlearn predict_proba runs forward-backward over the whole sample
Label switching after retraining
Too few transitions for credible inference
Controller churn
COUNTER-EVIDENCEWhat the literature says against it
Each finding is worded as the cited source reports it and links to that source.
LIMITATION
The reading: Library state probabilities can be traded as they come.
The hmmlearn posterior is computed from forward and backward passes over the whole sequence, so it uses later observations; the project README also states that the package is in limited-maintenance mode.
Evaluate current pretrained forecasting families such as Chronos-2 and TimesFM 2.5 alongside local models. Treat pretraining as a modeling prior and test financial usefulness independently.
Pretrained quantile forecasters, filed under the closest role (persistence in a series’ own history) because the vocabulary has no general forecasting role. Experimental: the cited 2026 benchmark finds gains over a random walk small and sparse.
Estimate calibrated forecast distributions or quantiles, then assess a separate cost-aware decision policy. This pinball loss is an evaluation option, not a claim about every model’s training loss.
01Signal inputs
Causal return, volatility or volume histories
Available covariates with known arrival times
Forecast quantiles and uncertainty
02Data specification
Audit pretraining dates and potential benchmark contamination
Align target transformations and forecast horizons
03Execution & infrastructure
Latency and resource budgets appropriate to horizon
Fallback behavior when forecasts are stale or unavailable
Weight license checked per checkpoint: TimesFM is Apache-2.0 through 2.5; the 3.0 weights are non-commercial
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITRolling origins
Evaluate zero-shot or fine-tuned forecasts at successive origins in calendar order.
2
LEAKAGEAudit the pretraining window
Check pretraining dates against the test period and pin the checkpoint, context length and preprocessing.
3
BASELINEBaselines to beat
A random walk and local models fitted on the same history.
4
METRICCalibration, then outcomes
Pinball loss and interval coverage first; net trading outcomes of a separate decision policy second.
5
REFUTATIONWhat refutes it
The pretrained model does not beat the random walk on returns, or its intervals miss their nominal coverage.
04Known failure modes
Good level forecasts with weak return predictability
Covariate leakage
Uncalibrated intervals
Computational overhead without economic benefit
COUNTER-EVIDENCEWhat the literature says against it
Each finding is worded as the cited source reports it and links to that source.
REPLICATION
The reading: Pretrained forecasters deliver tradable return forecasts.
A 2026 benchmark on five liquid US equities finds gains over the random-walk benchmark small and sparse: a one-sided Diebold–Mariano test rejects only for Chronos on AMZN and Moirai-2.0 on GOOG.
The reading: Any released checkpoint can be deployed.
The TimesFM repository keeps its code and its weights through version 2.5 under Apache-2.0 and restricts the 3.0 weights to non-commercial, non-production use.
Structure filings, announcements and other licensed text into testable event features. Separate the source event time, publication time, ingestion time and model completion time.
Language models turn timed documents into a return forecast that must clear a cost threshold. Contested on measurement: the cited point-in-time study states that a language model trained on later text has already seen the outcomes a backtest asks it to predict.
Measure incremental predictive value beyond structured event and price baselines, with a decision rule defined before the final holdout.
01Signal inputs
Original event documents and identifiers
Entity-linked structured features
Pre-event expectations where licensed
02Data specification
Versioned source documents and trustworthy availability times
Historical source coverage and revision tracking
03Execution & infrastructure
Bounded parsing and inference latency
Duplicate-event protection and late-event handling
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITSplit by event date
Train on earlier events and test on later ones; never split one event across folds.
2
LEAKAGEFreeze what the model knew
Pin model weights, prompts and document versions per experiment, and use the publication time, not the event time.
3
BASELINEBaselines to beat
Structured surprise and price features without any text.
4
STRESSDelay sensitivity
Shift the assumed publication time and re-run; add parsing and inference latency.
5
REFUTATIONWhat refutes it
Text adds nothing beyond the structured baselines on the final holdout, or the signal arrives after the price response.
04Known failure modes
Future knowledge embedded in model outputs
Text revisions or historical timestamp errors
Hallucinated facts
A signal arriving after the price response
COUNTER-EVIDENCEWhat the literature says against it
Each finding is worded as the cited source reports it and links to that source.
CRITIQUE
The reading: A language-model backtest measures a tradable signal.
Lee et al. state that language models used in financial backtests suffer from look-ahead bias, because a model trained on text published after the study period has already observed the outcomes it is asked to predict.
Research quotes that balance spread capture, fill probability, adverse selection and inventory. A broker-level bar backtest is insufficient for claims about a queue-sensitive strategy.
A closed-form stochastic-control solution that sets quotes from inventory and time to the horizon, on full-depth data. Active: the cited model is a stylized specification whose live value depends on venue-specific fills, queues, and adverse selection.
The stylized Avellaneda–Stoikov specification maximizes expected exponential utility of terminal wealth; production research must model venue-specific fills and controls.
01Signal inputs
Level 2 or Level 3 book events
Trades, inventory and tick sizes
Fill-intensity and short-horizon volatility estimates
02Data specification
Sequence-correct order-book replay
Venue fees, queue rules and latency measurements
03Execution & infrastructure
Rapid cancel/replace with acknowledgements
Inventory limits, disconnect cancellation and stale-book protection
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITOut-of-period calibration
Calibrate fill intensity and volatility on one period and replay another.
2
LEAKAGESequence-correct replay
Replay book events in exchange order with measured latency; a passive order fills only after the queue ahead of it.
3
BASELINEBaseline to beat
A symmetric fixed spread with the same inventory limits.
4
STRESSAdverse selection
Delayed cancellations, toxic flow and stale-book intervals.
5
REFUTATIONWhat refutes it
Inventory losses exceed spread capture once queue position and adverse selection are modeled, or results depend on fills the queue would not have given.
04Known failure modes
Optimistic passive fills
Incorrect queue position
Toxic order flow
Inventory loss dominating spread capture
IMPLEMENTATION STACK
NautilusTrader (L2/L3 order-book replay)hftbacktestRustFIX via QuickFIX
Compare TWAP, volume-aware schedules and constrained learned policies for a fixed parent-order mandate. Execution quality is evaluated separately from the signal that requested the trade.
A policy over child orders under a deadline, benchmarked against TWAP and volume schedules. Experimental for the learned policy, because the cited study reports its gains in simulated order books; the classical schedules it is compared with are the Active part.
Reduce implementation shortfall and completion risk under participation, inventory and deadline constraints; learned rewards must reflect the intended execution mandate.
01Signal inputs
Parent order and arrival price
Book state, volume and remaining inventory
Deadline and participation constraints
02Data specification
Quote/trade replay and realistic fill responses
Calibrated market impact and latency assumptions
03Execution & infrastructure
Hard limits external to the learned policy
Partial-fill reconciliation and deadline fallback
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITUnseen market periods
Train policies on one set of sessions and evaluate on later ones.
2
LEAKAGEOne cost model for all
Every policy receives identical parent orders, fill responses and impact assumptions.
3
BASELINEBaselines to beat
TWAP and a volume-aware schedule under the same participation and deadline limits.
4
METRICShortfall and completion
Implementation shortfall, its dispersion and the unfilled quantity, reported separately; ablate each reward term.
5
REFUTATIONWhat refutes it
The learned policy does not beat the schedules on unseen periods, or its gain comes from behavior the simulator rewards and a venue would not.
04Known failure modes
Simulator exploitation
Unmodeled impact
Reward hiding unfilled inventory
Strong simulation results failing to transfer
COUNTER-EVIDENCEWhat the literature says against it
Each finding is worded as the cited source reports it and links to that source.
LIMITATION
The reading: Learned execution policies beat schedule benchmarks.
Cheridito and Weiss report that their method outperforms traditional benchmark strategies in simulated limit order book environments; the result is a simulator result.
Translate forecasts into feasible positions using convex optimization, robust risk estimates and turnover control. Risk budgeting and hierarchical allocation provide useful alternative baselines.
A convex program solved from forecasts and risk estimates, so its data tier is model outputs. Active: the cited solver documentation settles the mathematics, and DeMiguel, Garlappi and Uppal report in a peer-reviewed study that no mean-variance variant they test consistently beats the equal-weight rule out of sample, so any gain depends on the quality of the inputs.
Balance forecast return, estimated risk and transaction costs under a defined mandate; compare with equal-weight, inverse-volatility and minimum-variance allocations.
01Signal inputs
Expected returns and covariance
Current positions and trading costs
Sector, factor and liquidity constraints
02Data specification
Consistent estimation horizon and units
Positive-semidefinite covariance and point-in-time exposures
03Execution & infrastructure
Convert weights into valid lots after optimization
Recheck cash, margin and exposure after rounding
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITForward rebalances
Estimate inputs on data through each rebalance date and hold until the next.
2
LEAKAGEOne horizon, one unit
Returns, covariance and costs share the same horizon and notional convention; exposures are point-in-time.
3
BASELINEBaselines to beat
Equal-weight, inverse-volatility and minimum-variance allocations.
4
STRESSPerturb the inputs
Shrink and perturb expected returns and covariance, tighten constraints, and round to tradable lots.
5
REFUTATIONWhat refutes it
Net of turnover the optimized book does not beat the three baselines, or small input changes move it to a different corner.
04Known failure modes
Unstable mean estimates driving corner allocations
Map bounded selection problems to QUBO or Ising form and compare QAOA, annealing or hybrid candidates with exact and heuristic classical solvers on the same objective.
A cardinality-constrained selection solved offline, filed under the slowest horizon because it is not an execution-time method, with classical solvers as the reference. Experimental: the cited IBM tutorial states that uniform random sampling performs about as well as its QAOA sampler at the 40-asset scale.
Find feasible, high-quality discrete portfolios and measure objective gaps, success probability and total runtime. A quantum formulation does not establish quantum advantage.
01Signal inputs
Frozen risk/return snapshot
Cardinality and allocation rules
Penalty scales and backend resource limits
02Data specification
Exact mapping tests on enumerable instances
Matched original and encoded objective units
03Execution & infrastructure
Separate offline optimization from order execution
Classical feasibility checks and deterministic fallback
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITFrozen snapshot
Every solver receives the same risk and return snapshot and the same encoded objective.
2
LEAKAGEVerify the mapping
Enumerate every state of a small instance and check that the encoded and the original objective agree.
3
BASELINEBaselines to beat
Exact enumeration where feasible, a mixed-integer solver, simulated annealing and uniform random feasible sampling.
4
METRICGap, success, total time
Objective gap, probability of reaching the optimum and wall-clock time including training, queues and post-processing; raw and post-processed results reported apart.
5
REFUTATIONWhat refutes it
A classical solver or random sampling reaches the same objective gap in less total time on the same instance.
04Known failure modes
Penalty scaling erasing objective resolution
Large encoded bit counts
Quantum time excluding training or queues
Post-processing receiving uncredited improvement
COUNTER-EVIDENCEWhat the literature says against it
Each finding is worded as the cited source reports it and links to that source.
LIMITATION
The reading: A QAOA sampler beats classical sampling on portfolio selection.
IBM’s multi-objective tutorial states that at its 40-asset scale uniform random sampling performs about as well as the QAOA sampler, and that it teaches a workflow rather than demonstrating quantum advantage.
Label each event by the first of three barriers its price path touches, then train a binary meta-model that decides whether to act on the primary signal and how much to stake.
A path-dependent labeling rule plus a binary act-or-abstain classifier that sits on any primary signal. Active: the cited methodology paper and the journal studies behind the official code rest on controlled experiments, not on independent live replication.
Fit a calibrated meta-probability with uniqueness-weighted log-loss under purged, embargoed folds. The primary model sets the side, the meta-model the size; barrier widths are research parameters.
LABEL SCHEMETriple-barrier label and meta-label [1]
1
Set three barriers at entry
Profit-taking and stop-loss levels are multiples of the volatility estimated at entry and need not be symmetric; the vertical barrier is a maximum holding period in bars. The multiplier notation is ours.
2
Walk the path to the first touch
The label depends on the whole path, so closing prices alone are not enough.
With the side fixed by the primary model, the meta-label is 1 when the bet closed with a positive return and 0 otherwise.
5
Weight by uniqueness
Labels that overlap in time share outcomes. Each is weighted by the average share of its interval that it does not share.
+1
Upper barrier first
The profit-taking level is reached before the stop and before the time-out.
−1
Lower barrier first
The stop-loss level is reached first.
0
Vertical barrier first
The holding period expires. The source also allows the sign of the return at expiry instead of zero.
Barrier widths, the holding period and the sizing map are research parameters chosen inside training folds.
01Signal inputs
Event times from a primary signal with a declared side
Price path from entry to the vertical barrier
Volatility estimate fixing barrier widths at entry
Entry-time features for the meta-model, including primary-model diagnostics
02Data specification
Bar or tick paths, not closing prices alone: the label is path-dependent
Label start and end time stored for every observation
Volatility and features computed from information available at entry
03Execution & infrastructure
Live profit-take, stop and time-out rules identical to those used for labeling
Sizing map with hard caps applied outside the model
Recalibration schedule and drift monitor for meta-probabilities
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITSplit by label interval
Order events by entry time and build folds from whole label intervals, from entry to first touch.
2
LEAKAGEPurge and embargo
Drop training labels that overlap a test label, and those that start within an embargo after it. The cited paper states that an embargo near 1% of the sample often suffices; skfolio sets both sizes to zero by default, so set them.
3
BASELINEBaselines to beat
The primary signal acted on every time at fixed size, and a fixed-horizon label with the same classifier.
4
METRICClassifier first, calibration second
Precision, recall and F1 of the meta-model, then Brier score and a reliability curve before any sizing map is applied.
5
STRESSA distribution of paths
Combinatorial purged cross-validation yields several backtest paths instead of one. Report their spread and deflate the best result by the number of barrier and horizon settings tried.
6
REFUTATIONWhat refutes it
Out of fold, the meta-model does not raise precision over the always-act baseline, or sized positions do no better than fixed size once probabilities are calibrated.
04Known failure modes
Overlapping labels treated as independent samples
k-fold leakage across label intervals
Meta-model asked to rescue a primary signal with no recall: it filters bets, it does not propose them
Uncalibrated probabilities driving position size
Barrier widths tuned on the evaluation period
COUNTER-EVIDENCEWhat the literature says against it
Each finding is worded as the cited source reports it and links to that source.
LIMITATION
The reading: A meta-model can rescue a weak primary signal.
López de Prado describes the secondary model as one that does not propose bets: it decides whether a positive from the primary model is true or false, so it cannot add recall the primary model lacks.
The reading: Calibrated probabilities always improve position sizing.
Meyer, Barziy and Joubert, whose abstract the official repository reproduces, report that calibration significantly improves fixed sizing methods, while methods that estimate their sizing function from training data gain no significant advantage.
The reading: A purged scheme removes validation leakage at no cost.
A 2026 preprint proves that training sufficiency, test coverage and temporal causality cannot all hold in one validation scheme, and bounds the leakage bias that results.
NEURAL MICROSTRUCTURETens to hundreds of book events
/12
Deep order-book classification
Classify the next smoothed mid-price move as down, stationary or up from a tensor of recent order-book states. Published accuracy transfers poorly, so the label and the out-of-period test come first.
A convolution, inception and LSTM network that classifies the next mid-price move from full-depth book states. Contested: the cited LOBCAST re-run measures F1 of 48–61% for fifteen published models on 2021–2022 Nasdaq data, well below their benchmark scores.
Minimize three-class cross-entropy for a convolution, inception and LSTM network on a 100 × 40 book window. Label accuracy is a research metric; the trading rule is tested separately, net of costs.
ARCHITECTUREAs the cited paper states it
Layers and dimensions are the ones the source prints. Where a paper does not state a size, none is shown. Notation: a×b@c is a convolution with c filters of size a×b. The first index runs over time, the second over book features.
The 100 most recent book states. Each holds price and size for ten ask and ten bid levels.
02
Convolution block 1
The strided filter pairs price with size at each level. Leaky ReLU with slope 0.01; zero padding keeps the time dimension.
03
Convolution block 2100 × 10
A second strided filter combines the two sides of each level, in the manner of a micro-price.
04
Convolution block 3100 × 1
One 1×10 filter integrates the ten levels.
05
Inception module
Three parallel paths over time with 32 filters in every convolution, concatenated.
06
LSTM64 units
64 units in place of a fully connected layer. The paper counts about 60,000 parameters for the whole network, about ten times fewer than it would have with one fully connected layer of 64 units in this place.
07
Softmax3 units
Probabilities of down, stationary and up.
The paper states the feature-map sizes 100 × 10 and 100 × 1; the other intermediate sizes are not printed and are not shown here.
TRAINING, AS PUBLISHED
Categorical cross-entropy
Adam with learning rate 0.01 and epsilon 1
Mini-batches of 32
Training stops when validation accuracy has not improved for 20 epochs
The average of the best ask and the best bid at each book event.
2
Average before and after
One mean over the previous k mid-prices and one over the next k. The paper’s printed equation for m₋(t) runs its sum from i = 0 to k over 1/k, which is k + 1 terms, while its prose and the lab above use a k-term mean. Fix one convention and version it.
3
Take the relative change
Smoothing both sides gives more consistent labels than comparing with the current price alone.
4
Threshold into three classes
The threshold sets the class balance, so it is versioned with the dataset.
+1
Up
The smoothed change exceeds the threshold.
0
Stationary
The smoothed change stays within the threshold.
−1
Down
The smoothed change is below the negative threshold.
The FI-2010 benchmark smooths only the future window and fixes the threshold at 0.002. LOBCAST reports that its class mix moves with the horizon: 18/63/19% up, stationary and down at k = 1, and 37/25/38% at k = 10.
01Signal inputs
Ten ask and ten bid levels of price and size per book event (40 features)
Window of the 100 most recent book states
Smoothed mid-price label with horizon k and threshold α
02Data specification
Sequence-correct L2 or L3 replay per venue; months of history, not days
Rolling normalization fitted on past sessions only, per instrument
Label horizon, threshold and class balance versioned with the dataset
03Execution & infrastructure
Feature and inference latency measured against the label horizon
Decision rule stated net of spread, fees and queue position
Stale-book and feed-gap handling outside the classifier
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITChronological split per venue
Train, validation and test in calendar order. DeepLOB’s venue study uses six months, then three, then three; FI-2010 uses nine anchored day-forward folds.
2
LEAKAGENormalize with the past only
DeepLOB standardizes each day with the statistics of the previous five days. Windows that straddle a split boundary are dropped.
3
BASELINEBaselines to beat
A logistic model on order-flow imbalance and the majority class, with several seeds for every model.
4
METRICClass-aware metrics
Macro F1, per-class precision and recall, and the confusion matrix. Accuracy alone misleads when one class dominates.
5
STRESSTransfer test
Unseen instruments and a later period, with the threshold re-set against the spread.
6
REFUTATIONWhat refutes it
F1 falls toward the baseline on the later period or on unseen instruments, or the predicted moves stay inside the spread after costs.
04Known failure modes
Benchmark scores that do not transfer: LOBCAST measures F1 of 48–61% for fifteen published models on 2021–2022 Nasdaq data
Label artifacts: smoothing tied to horizon, class balance moving with k
Predicted moves smaller than the spread
Normalization or sampling that uses later sessions
Hyperparameter sensitivity mistaken for an architecture gain
COUNTER-EVIDENCEWhat the literature says against it
Each finding is worded as the cited source reports it and links to that source.
REPLICATION
The reading: Benchmark accuracy carries over to current markets.
LOBCAST re-ran fifteen published models. On its 2021 and 2022 Nasdaq datasets F1 ranged from 48% to 61%, significantly lower than on FI-2010, and about half of the hyperparameter-search runs diverged (F1 of 33% or less).
The reading: Deeper architectures explain the reported gains.
TLOB’s authors report that a simple MLP adaptation surpasses earlier state-of-the-art models and that predictability has declined over time, by 6.68 points of F1.
ORDER-FLOW CLASSIFICATIONVolume time: buckets to sessions
/13
Order-flow classification & toxicity
Sign trades with the tick rule, a quote rule or bulk volume classification, then measure imbalance per volume bucket (VPIN) or at the best quotes (OFI). VPIN’s predictive claims are disputed.
Rule-based trade signing and imbalance measures on trades and best quotes; the OFI part is a linear regression. Contested: among the cited papers, Andersen and Bondarenko dispute the predictive content of VPIN, Chakrabarty, Pascual and Shkilko the accuracy of bulk classification, and the method’s authors have published a reply.
Estimate signed flow with a stated classifier and test whether its imbalance adds information on volatility or price change beyond volume and realized volatility. No threshold is read as a crash predictor.
No sign per trade. The volume of each time bar is split into buy and sell shares by the standardized price change over the bar; the original uses one-minute bars.
+1
Buyer-initiated
The aggressor bought: the print is above the midpoint, or on an uptick.
−1
Seller-initiated
The aggressor sold: the print is below the midpoint, or on a downtick.
Lee and Ready document that quotes may be recorded ahead of the trades that triggered them and that trades inside the spread are not readily classifiable.
01Signal inputs
Trade prints with prices and sizes
Best bid and ask prices and sizes, sequenced with the trades
Aggressor flag where the venue publishes one
Volume-bucket size and bar length
02Data specification
Trade and quote streams on one clock with documented timestamp precision
Venue rules for hidden, auction and off-book prints
Exchange aggressor flags or order-matched data as ground truth for a sample
03Execution & infrastructure
Streaming bucket updates with late and corrected print handling
Feature freshness and feed-gap flags
Use as a risk or sizing input with independent limits
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITHold out sessions
Fix bucket size, bar length and any regression on earlier sessions and evaluate on later ones.
2
LEAKAGEOne clock
Align trades and quotes on exchange timestamps and sign each print against the quote that prevailed before it.
3
BASELINEBaselines to beat
Volume and realized volatility alone for volatility prediction; the tick rule for trade signing.
4
METRICClassifier and measure, separately
Signing accuracy against true aggressor flags by rule and liquidity tier, then the incremental fit of imbalance for price change or volatility.
5
STRESSSwap the classifier
Recompute VPIN under the tick rule, Lee–Ready and bulk classification and report where the readings disagree.
6
REFUTATIONWhat refutes it
The imbalance measure adds nothing once volume and realized volatility are in the regression, or its behavior reverses when the classification rule changes.
04Known failure modes
Trade signs that disagree across classification rules
A toxicity reading that restates volume and volatility: Andersen & Bondarenko find no incremental power once both are controlled
Timestamp misalignment between trades and quotes
Bucket and bar choices tuned to one episode
Imbalance measures that react after the move
COUNTER-EVIDENCEWhat the literature says against it
Each finding is worded as the cited source reports it and links to that source.
CRITIQUE
The reading: VPIN warns of toxicity-induced volatility in advance.
Andersen and Bondarenko find that VPIN “is a poor predictor of short run volatility, that it did not reach an all-time high prior, but rather after, the flash crash, and that its predictive content is due primarily to a mechanical relation with the underlying trading intensity”.
The reading: VPIN adds information beyond volume and volatility.
After controlling for current volume and volatility, Andersen and Bondarenko find no incremental predictive power, and note that its behavior reverses under more standard classification.
The reading: Bulk volume classification signs flow accurately.
Chakrabarty, Pascual and Shkilko find, for a large sample of equities, that the tick rule and Lee–Ready classify significantly better than bulk volume classification and yield more accurate VPIN estimates.
Easley, López de Prado and O’Hara answer that tick rules and bulk classification are both reasonable classifiers of the aggressor side, and that bulk classification is better linked to proxies of information-based trading.
Forecast several horizons at once with attention over time patches or over variates, emitting quantiles. A one-layer linear model is a mandatory competitor: it has beaten published Transformers.
An attention forecaster trained from scratch, filed under the closest role (persistence in a series’ own history) because the vocabulary has no general forecasting role. Experimental on returns: the cited studies report linear models ahead of Transformers on nine benchmarks and a mean directional accuracy of 50.08% across 918 runs.
Minimize pinball loss summed over horizons and quantiles; accept the attention model only if it beats the linear family and a last-value forecast on identical rolling origins.
ARCHITECTUREAs the cited paper states it
Layers and dimensions are the ones the source prints. Where a paper does not state a size, none is shown.
Each variate is processed as its own univariate series; all channels share one backbone.
02
Instance normalization
Each channel is normalized before it is cut into patches.
03
PatchingN = 42 patches
Patching cuts the number of input tokens from L to about L/S.
04
Projection and positionD = 128 per patch
A trainable linear projection maps each patch to the latent dimension; a learnable position encoding is added.
05
Transformer encoder
A vanilla encoder with scaled dot-product attention over the patches.
06
Flatten and linear head
The encoder output is flattened and mapped to the forecast horizon.
07
ForecastT steps per channel
Trained on mean squared error, averaged over channels.
The Temporal Fusion Transformer, the other cited architecture, combines static covariate encoders, variable selection networks, gated residual networks, a sequence-to-sequence LSTM layer and interpretable multi-head attention, and is trained on a joint quantile loss.
A single linear map along the time axis, shared across variates.
03
ForecastT steps per variate
All horizons are produced in one step.
DLinear first separates a moving-average trend from the remainder; NLinear subtracts the last value and adds it back.
01Signal inputs
Look-back window of targets and observed covariates
Known-in-advance inputs with their availability times
Static instrument metadata
Forecast horizons and quantile levels
02Data specification
Inputs classified as static, observed-only or known-in-advance before modeling
Targets defined on returns, volatility or volume; price-level results reported separately
Per-series normalization fitted on the look-back or training span only
03Execution & infrastructure
Inference budget set by the shortest horizon
Monotone quantile post-processing and a stale-input fallback
Model revision, look-back and patch settings pinned per experiment
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITRolling origins
Evaluate at successive forecast origins in calendar order, with the same origins for every model.
2
LEAKAGEClassify the inputs first
Mark each input as static, observed-only or known in advance, and normalize each series on its look-back only.
3
BASELINEBaselines to beat
The last value, the LTSF-Linear family and a tuned tree model, with the same tuning budget.
4
METRICError and decision, apart
Pinball loss and coverage per horizon, Diebold–Mariano tests against the random walk, then directional accuracy and net decision value.
5
STRESSSeeds and budgets
Several seeds per architecture and equal hyperparameter budgets, so that a ranking is not a tuning artifact.
6
REFUTATIONWhat refutes it
The linear family or the last-value forecast matches the attention model within seed dispersion, or directional accuracy stays near one half.
04Known failure modes
Linear or last-value baselines matching the Transformer: Zeng et al. report one-layer linear models ahead on nine benchmarks
Error rankings that leave directional accuracy near one half
Known-in-advance inputs that were not actually known
Electricity or traffic benchmarks assumed to transfer to returns
COUNTER-EVIDENCEWhat the literature says against it
Each finding is worded as the cited source reports it and links to that source.
CRITIQUE
The reading: Transformers are the strongest long-horizon forecasters.
Zeng et al. report that one-layer linear models outperform the Transformer forecasters they compare on nine benchmarks, surpassing the strongest of them in most cases by 20% to 50%, and that repeating the last value beats all of them on the Exchange-Rate dataset by around 45%.
The reading: Neural forecasters beat a random walk on equity returns.
A 2026 benchmark on five liquid US equities, which includes PatchTST and iTransformer, finds gains over the random-walk benchmark small and sparse: a one-sided Diebold–Mariano test rejects only for Chronos on AMZN and Moirai-2.0 on GOOG.
The reading: A lower forecast error means better direction calls.
A 2026 controlled comparison of nine architectures over 918 runs ranks ModernTCN and PatchTST first and second by error, while mean directional accuracy is 50.08%.
The reading: A longer look-back always helps attention.
The iTransformer authors state that Transformers with temporal tokens suffer performance degradation and computation explosion with larger look-back windows.
Forecast realized variance from its daily, weekly and monthly averages, keep GARCH(1,1) as the return-based reference, and admit a neural extension only where it beats a well-estimated linear model.
A linear autoregression of realized variance on its daily, weekly and monthly averages, built from intraday bars. Established for the HAR baseline, which the cited HARNet and 2026 portfolio studies take as their reference; the neural extensions are Active, because the HARNet authors report little to no gain over a well-estimated HAR.
Estimate the heterogeneous autoregression by OLS, WLS or on logarithms and rank forecasts by QLIKE and squared error; neural models start from, and are judged against, the fitted linear model.
ARCHITECTUREAs the cited paper states it
Layers and dimensions are the ones the source prints. Where a paper does not state a size, none is shown.
The receptive field of the network. Inputs are scaled linearly before they enter the model.
02
Layer 1, daily
With a first aggregation period of one day, the paper omits this convolution and passes the series through.
03
Layer 2, weekly
Initialized as a five-day average.
04
Layer 3, monthly
Initialized as an average of four weekly values, twenty days in all.
05
Linear output
A weighted sum of the three layer outputs plus an intercept, initialized at the fitted HAR coefficients. Before training, the network reproduces the HAR forecast exactly.
06
One-day-ahead forecast13 trainable parameters
The forecast is clipped from below so that it stays positive.
The paper also defines a 19-parameter variant with aggregation periods (1, 5, 20, 40, 80). HARNet uses 20 days for the monthly component where the card formula uses 22.
TRAINING, AS PUBLISHED
10,000 Adam iterations at learning rate 0.0001
Batch size 4, five consecutive labels per sample
MAE, MSE and QLIKE compared as training losses
Four years of training data and the following year as test, for 15 pairs
Yesterday’s realized variance and its 5-day and 22-day averages.
02
Linear regression4 coefficients
Estimated by OLS, by WLS or on logarithms.
03
One-day-ahead forecast
The reference every neural variant is initialized at and judged against.
01Signal inputs
Intraday returns for the realized measure
Daily, weekly and monthly averages of realized variance
Daily returns for the GARCH reference
Optional intraday realized measures and a market-volatility proxy
02Data specification
Cleaned intraday prices with a documented sampling frequency
Overnight and auction treatment fixed in advance
Realized measure definition versioned: variance, bipower or kernel
03Execution & infrastructure
Forecast available before the sizing decision of the next session
Floor and cap on the forecast used for leverage
Fallback to the last valid forecast on data gaps
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITRolling or expanding re-estimation
Re-fit on a moving window and forecast the next period. The HARNet study trains on four years and tests on the following year, fifteen times.
2
LEAKAGEFix the measure
Sampling frequency, overnight treatment and the realized-measure definition are set before the first fit.
3
BASELINEBaselines to beat
HAR estimated by OLS, by WLS and on logarithms; GARCH(1,1) and a leverage-aware variant as return-based references.
4
METRICQLIKE and squared error
Compared with Diebold–Mariano or model-confidence-set tests.
5
STRESSCarry the forecast into sizing
Measure the effect on position size, turnover and realized risk, not only on forecast error.
6
REFUTATIONWhat refutes it
The neural model does not beat HAR fitted by WLS or on logarithms, or its lower forecast error does not change realized portfolio risk.
04Known failure modes
Neural gain that disappears against a robustly estimated HAR: HARNet’s authors report little to no improvement over WLS and log-OLS fits
Microstructure noise or sparse sampling in the realized measure
Lowest forecast error not producing the best portfolio
Leverage inflated after calm periods
Overnight gaps excluded from the measure but present in the risk
COUNTER-EVIDENCEWhat the literature says against it
Each finding is worded as the cited source reports it and links to that source.
LIMITATION
The reading: A neural network improves on HAR.
The HARNet authors report substantial improvements over a HAR baseline fitted by ordinary least squares, but often little to no improvement when the baseline is fitted by weighted least squares or on logarithms.
The reading: A more complex volatility model beats a simple one.
Hansen and Lunde compare 330 ARCH-type models and find no evidence that GARCH(1,1) is outperformed for exchange rates, while it is clearly inferior to models with a leverage effect for IBM returns.
The reading: The most accurate forecast gives the best portfolio.
A 2026 study of weekly realized volatility for 465 S&P 500 equities reports that the model with the lowest forecast error, the one with the best cross-sectional ranking and the one with the highest portfolio Sharpe ratio are three different models.
Wrap any forecaster in prediction sets with controlled long-run miscoverage, calibrate classifier probabilities, and size from the result. Coverage is a guarantee; profitability is tested apart.
A calibration layer on model outputs that sizes or abstains, at whatever horizon the forecaster uses. The coverage bound is proved in the cited paper by Gibbs and Candès, but as a sizing method it is Experimental: the cited pre-registered study reports 0.745 coverage against a 0.750 target and still ranked last of eleven books on Sharpe.
Hold long-run miscoverage at the target under distribution shift by adapting the working level online; calibrate probabilities on held-out data; then map uncertainty to a capped position.
01Signal inputs
Forecasts or class probabilities from any model
A calibration stream disjoint from the training data
Target miscoverage and adaptation step
Risk caps for the sizing map
02Data specification
Calibration data later than training data and earlier than deployment
Outcome arrival times: the update needs the realized value
Enough calibration points per regime to inspect conditional coverage
03Execution & infrastructure
Infinite or empty sets handled explicitly as abstentions
Sizing caps enforced outside the calibration layer
Recalibration when coverage drifts beyond a tolerance
EVALUATION PROTOCOLFrom the split to the result that would refute it
1
SPLITThree disjoint spans
Training data, then a later calibration stream, then deployment. Nothing used to fit the model calibrates it.
2
LEAKAGERespect outcome arrival
The online update uses a realized value only once it is known.
3
BASELINEBaselines to beat
A fixed-level split-conformal interval for coverage; volatility-scaled and equal-weight books for the economics.
4
METRICCoverage, width, calibration
Rolling coverage against the target, interval width against realized error, and a reliability diagram with ECE and Brier score before and after calibration.
5
STRESSCondition on stress
Report coverage inside high-volatility periods, not only in aggregate.
6
REFUTATIONWhat refutes it
Long-run coverage misses the target, coverage collapses in stress states, or the sized book does not beat a volatility-scaled book under a protocol registered in advance.
04Known failure modes
Marginal coverage read as conditional coverage
Aggregate coverage hiding under-coverage in stress
Exchangeability assumed for drifting markets
Calibration fitted on data the model has seen
Valid intervals with no economic value: a pre-registered 2026 study reports 0.745 coverage against a 0.750 target and a last place among eleven books on Sharpe
COUNTER-EVIDENCEWhat the literature says against it
Each finding is worded as the cited source reports it and links to that source.
REPLICATION
The reading: Valid coverage makes a sizing rule profitable.
A pre-registered 2026 study that sizes eight US-listed ETFs from conformal interval width reports realized coverage of 0.745 against a 0.750 target on its sealed 2022–2024 window, while both registered configurations ranked last of eleven entries on Sharpe and Calmar.
Han and Lin find that recent-only calibration of next-day range forecasts is near nominal in aggregate but undercovers substantially during high-volatility states.
The reading: A conformal interval is valid for each individual forecast.
Angelopoulos and Bates state that conformal prediction guarantees marginal coverage and that conditional coverage is impossible to achieve in the most general case.