AstrolabeMarket TerminalBETA

Do market regimes help short-horizon direction models?

What the CP2 research found for gold (XAUUSD) and a Nasdaq-100 ETF (QQQM) on 5-minute bars, with every result labelled by how it was checked.

  • CP2 CAPSTONE · SUNWAY UNIVERSITY
  • RESULTS EXPORT 2026-11-05
  • EXPLORATORY, NOT CONFIRMATORY
  • NO TRADING CLAIM

Evidence labels used here: reported stated by the run owner; independently tested re-checked by a tool, with method, date and scope; source-inspected read from code or committed files.

The question

Markets move through spells that look different: calm drifts, choppy ranges, violent swings. The study asks whether a model that first recognises which spell the market is in, and then uses a classifier trained for that spell, predicts short-horizon direction better than one classifier trained on everything.

The answer matters less as a trading edge than as a test of method: can the terminal show, honestly and repeatably, where such a model helps, where it does not, and how much of that can be trusted?

Method

Data

M5 bars for XAUUSD (gold) and QQQM (a Nasdaq-100 ETF), with D1/H4/H1/M15/M5 context aligned to the M5 observation coordinate. Prices stay on the owner's computer and are not published.

Source: experiments/cp2_report_draft2_v1/manuscript_v1/CH03/CH03_DRAFT2.md §3.2.1 · source-inspected

Regimes

A Gaussian hidden Markov model with K4 states and diag covariance assigns each bar a latent state; the model is frozen after training.

Source: experiments/cp2_report_draft2_v1/manuscript_v1/CH03/CH03_DRAFT2.md §3.2.4 Table 3.3 · source-inspected

Target

Down, Neutral or Up from the close return to the sixth subsequent M5 row: Up when r_t>0.001, Down when r_t<−0.001, Neutral in between, with an elapsed horizon of 25–35 minutes.

Source: experiments/cp2_report_draft2_v1/manuscript_v1/CH03/CH03_DRAFT2.md §3.2.2 · source-inspected

Models

Adaptive: an eligible state-specific XGBoost bundle chosen by the regime model, with a global fallback. Global: one XGBoost classifier that pools the original feature population.

Source: experiments/cp2_report_draft2_v1/manuscript_v1/CH03/CH03_DRAFT2.md §3.2.4 · source-inspected

Evaluation

Four expanding walk-forward development folds, a fixed holdout and a later historical fresh period, each reported separately. Accuracy, fixed-three-class Macro-F1, support-weighted F1 and confusion matrices describe complementary aspects of classification.

Source: experiments/cp2_report_draft2_v1/manuscript_v1/CH03/CH03_DRAFT2.md §3.2.6, §3.4.1 · source-inspected

Results at a glance

Results at a glance, read from the results export
FindingResultEvidence
Regime decoding check (MR-01)CHECKReplication gate pass in every population. Causal and whole-sequence labels agree on 0.9861179970252851 to 0.9985443959243085 of rows. Largest change in Macro-F1: not in the results export.reported (owner-run; metrics file as posted; not recomputed)
Adaptive vs globalCOMPARISONWhole-sequence Macro-F1 of the regime-routed model next to the relisted global model for each instrument's fixed holdout and historical fresh period, as the claims ledger records them (chart below). Count of populations in which it is ahead: not in the results export.reported
TimesFM 2.5, zero-shotNEGATIVEMean absolute error, model minus persistence (above 0 = model worse): holdout 9.049291580353995e-05, fresh 0.00015244070356988247.reported · containment_qualified: false
EDL studyEXPLORATORYK01: positive exploratory diagnostic; K02–K10, K16–K22: inconclusive. Development period only; not confirmed and not a claim.reported · not confirmatory
Synthetic regime recoverySYNTHETICNumber of synthetic fits: 1920. Number of parameter cells: not in the results export. They verify the method; they say nothing about real markets.reported (owner-run on SACF; file as delivered; not re-fitted)

Each result is read at build time from committed result files, with its source and evidence label. Negative and inconclusive results are shown as found.

Every value in the results export, with source and evidence

Headline model figures: development, fixed holdout and historical fresh

  • values as written in the Chapter 4 claims ledger
  • populations reported separately, never pooled

Direction-model figures exactly as the Chapter 4 results-claims ledger of the CP2 report draft records them: the regime-adaptive model under its original whole-sequence regime labels, and the unchanged global model relisted from its saved metrics, for each instrument on the development folds (equal mean of the four folds, as the file gives it), the fixed holdout and the historical fresh period. Populations are named as the ledger names them; the ledger rows give no calendar dates. Exploratory research results, not a trading recommendation and not investment advice.

Headline model figures: development, fixed holdout and historical fresh: values with source and evidence label
ValueResultEvidenceSource
QQQM development (equal mean of the four folds) whole-sequence Macro-F1 (C04-R179)0.3896784170912872reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R179/value_as_written
QQQM development (equal mean of the four folds) whole-sequence accuracy (C04-R181)0.468969167659124reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R181/value_as_written
QQQM fixed holdout whole-sequence Macro-F1 (C04-R99)0.391185042763235reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R99/value_as_written
QQQM fixed holdout whole-sequence accuracy (C04-R101)0.47222013946164804reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R101/value_as_written
QQQM fixed holdout global Macro-F1 (relisted) (C04-R105)0.3559988734227451reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R105/value_as_written
QQQM historical fresh whole-sequence Macro-F1 (C04-R109)0.3943384106077888reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R109/value_as_written
QQQM historical fresh whole-sequence accuracy (C04-R111)0.44534412955465585reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R111/value_as_written
QQQM historical fresh global Macro-F1 (relisted) (C04-R115)0.3624472129438476reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R115/value_as_written
XAUUSD development (equal mean of the four folds) whole-sequence Macro-F1 (C04-R185)0.3439477396205161reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R185/value_as_written
XAUUSD development (equal mean of the four folds) whole-sequence accuracy (C04-R187)0.7081746552856205reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R187/value_as_written
XAUUSD fixed holdout whole-sequence Macro-F1 (C04-R159)0.40280872264291845reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R159/value_as_written
XAUUSD fixed holdout whole-sequence accuracy (C04-R161)0.4931092987260388reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R161/value_as_written
XAUUSD fixed holdout global Macro-F1 (relisted) (C04-R165)0.3634154387733373reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R165/value_as_written
XAUUSD historical fresh whole-sequence Macro-F1 (C04-R169)0.37821750685585914reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R169/value_as_written
XAUUSD historical fresh whole-sequence accuracy (C04-R171)0.4774417451660882reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R171/value_as_written
XAUUSD historical fresh global Macro-F1 (relisted) (C04-R175)0.3465026681071915reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R175/value_as_written

MR-01: causal versus whole-sequence regime labels

REPLICATION GATE PASS IN EVERY POPULATION
  • sensitivity analysis of a stated limitation; exploratory; not a new model, not confirmatory
  • primary variant: causal_online_viterbi_cold

The original regime labels were decoded over whole sequences, which can use bars after the row being labelled. MR-01 re-decoded them causally (online, no later bars) and compared the two. For each instrument and population the table gives the replication-gate status and the share of rows on which the causal labels equal the whole-sequence labels, as written in the metrics file. Populations are not pooled and instruments are not averaged.

Dot plot of MR-01 label agreement for 12 populations (QQQM and XAUUSD development folds 1 to 4, fixed holdout, historical fresh); every value is between 0.98 and 1.
Each dot is one population's share of rows on which causal (online) regime labels equal the whole-sequence labels, as written in the MR-01 metrics file. The lowest plotted value is 0.9861179970252851 (XAUUSD · historical fresh) and the highest is 0.9985443959243085 (QQQM · development fold 4). The axis starts at 0.98, not 0. Evidence label: reported (owner-run; metrics file as posted; not recomputed).
MR-01: causal versus whole-sequence regime labels: values with source and evidence label
ValueResultEvidenceSource
Label recorded with the resultsensitivity analysis of a stated limitation; exploratory; not a new model, not confirmatoryreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /label
Primary causal variantcausal_online_viterbi_coldreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /primary_variant
Run date2026-11-02reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /run_date
QQQM development fold 1: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_1/status
QQQM development fold 1: label agreement, causal vs whole-sequence0.9958978430594151reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_1/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
QQQM development fold 2: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_2/status
QQQM development fold 2: label agreement, causal vs whole-sequence0.9943099113404791reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_2/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
QQQM development fold 3: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_3/status
QQQM development fold 3: label agreement, causal vs whole-sequence0.9915310308323408reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_3/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
QQQM development fold 4: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_4/status
QQQM development fold 4: label agreement, causal vs whole-sequence0.9985443959243085reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_4/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
QQQM fixed holdout: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1fixed_holdout/status
QQQM fixed holdout: label agreement, causal vs whole-sequence0.9939266701657045reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1fixed_holdout/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
QQQM historical fresh: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1historical_fresh/status
QQQM historical fresh: label agreement, causal vs whole-sequence0.9919028340080972reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1historical_fresh/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
XAUUSD development fold 1: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_1/status
XAUUSD development fold 1: label agreement, causal vs whole-sequence0.9983453709783322reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_1/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
XAUUSD development fold 2: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_2/status
XAUUSD development fold 2: label agreement, causal vs whole-sequence0.9965331582403152reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_2/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
XAUUSD development fold 3: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_3/status
XAUUSD development fold 3: label agreement, causal vs whole-sequence0.9977675640183847reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_3/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
XAUUSD development fold 4: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_4/status
XAUUSD development fold 4: label agreement, causal vs whole-sequence0.9962967826657912reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_4/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
XAUUSD fixed holdout: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1fixed_holdout/status
XAUUSD fixed holdout: label agreement, causal vs whole-sequence0.9903559947612811reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1fixed_holdout/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
XAUUSD historical fresh: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1historical_fresh/status
XAUUSD historical fresh: label agreement, causal vs whole-sequence0.9861179970252851reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1historical_fresh/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence

SYNTH: regime recovery on synthetic series

  • synthetic method verification only; says nothing about real markets

Synthetic method verification only; says nothing about real markets. The CP2 hidden-Markov configuration was fitted to simulated series whose true regimes are known, over a grid of regime counts, persistence, separation and series length. Recovery is the adjusted Rand index (ARI) between true and fitted states, where 1 is exact. The chart places every grid cell's stored mean ARI by regime separation, for whole-sequence and causal decoding.

Strip chart of synthetic regime-recovery ARI for every parameter cell, grouped by the four regime separations, whole-sequence and causal decoding side by side. Synthetic method verification only.
Synthetic method verification only; says nothing about real markets. Each mark is one parameter cell's stored mean adjusted Rand index over its seeds. Lowest to highest plotted cell value, as written: separation 0.25: whole-sequence 0.4851819766130164 to 0.9519954448355454, causal 0.3524916247961309 to 0.8977456933609339 (48 cells each); separation 0.5: whole-sequence 0.6980329061003921 to 0.9981760999204768, causal 0.6632255106932002 to 0.9877490547006588 (48 cells each); separation 1.0: whole-sequence 0.7598777380462557 to 0.9998811363963032, causal 0.753977393185088 to 0.9994108029282808 (48 cells each); separation 2.0: whole-sequence 0.7552420681638202 to 0.9999446847527047, causal 0.7552420681638202 to 0.9999446847527047 (48 cells each). Evidence label: reported (owner-run on SACF; file as delivered; not re-fitted).
SYNTH: regime recovery on synthetic series: values with source and evidence label
ValueResultEvidenceSource
Label recorded with the resultsynthetic method verification only; says nothing about real marketsreported (owner-run on SACF; file as delivered; not re-fitted)experiments/cp2_synth_regime_v1/results/run1/synth_regime_metrics_v1.json @ fa88153ed383 · /label
Number of fits1920reported (owner-run on SACF; file as delivered; not re-fitted)experiments/cp2_synth_regime_v1/results/run1/synth_regime_metrics_v1.json @ fa88153ed383 · /n_fits
Fitted model: number of states4reported (owner-run on SACF; file as delivered; not re-fitted)experiments/cp2_synth_regime_v1/results/run1/synth_regime_metrics_v1.json @ fa88153ed383 · /fitted_config/n_states
Fitted model: covariance typediagreported (owner-run on SACF; file as delivered; not re-fitted)experiments/cp2_synth_regime_v1/results/run1/synth_regime_metrics_v1.json @ fa88153ed383 · /fitted_config/covariance_type

POWER: design sensitivity of the TimesFM and EDL results

  • design-sensitivity analysis from committed summaries; approximate; not a re-analysis of the data

How small an effect the committed TimesFM and EDL designs could have detected, computed from their committed summaries. It does not re-test any hypothesis, does not change a committed result and does not say that a result would change with more data; sessions-needed figures are planning magnitudes only.

POWER: design sensitivity of the TimesFM and EDL results: values with source and evidence label
ValueResultEvidenceSource
Label recorded with the resultdesign-sensitivity analysis from committed summaries; approximate; not a re-analysis of the dataindependently computed from committed summaries (method experiments/cp2_power_v1/METHOD.md); inputs reportedexperiments/cp2_power_v1/results/power_metrics.json @ fa88153ed383 · /label
TimesFM holdout: sessions185independently computed from committed summaries (method experiments/cp2_power_v1/METHOD.md); inputs reportedexperiments/cp2_power_v1/results/power_metrics.json @ fa88153ed383 · /timesfm/regions/1/n_sessions
TimesFM holdout: observed directionmodel_worseindependently computed from committed summaries (method experiments/cp2_power_v1/METHOD.md); inputs reportedexperiments/cp2_power_v1/results/power_metrics.json @ fa88153ed383 · /timesfm/regions/1/observed_direction
TimesFM holdout: minimum detectable improvement at 80% power, as a fraction of persistence MAE0.021622812464768215independently computed from committed summaries (method experiments/cp2_power_v1/METHOD.md); inputs reportedexperiments/cp2_power_v1/results/power_metrics.json @ fa88153ed383 · /timesfm/regions/1/sensitivity/mde_fraction_of_mae_persistence/power_0.80
TimesFM fresh: sessions15independently computed from committed summaries (method experiments/cp2_power_v1/METHOD.md); inputs reportedexperiments/cp2_power_v1/results/power_metrics.json @ fa88153ed383 · /timesfm/regions/0/n_sessions
TimesFM fresh: observed directionmodel_worseindependently computed from committed summaries (method experiments/cp2_power_v1/METHOD.md); inputs reportedexperiments/cp2_power_v1/results/power_metrics.json @ fa88153ed383 · /timesfm/regions/0/observed_direction
TimesFM fresh: minimum detectable improvement at 80% power, as a fraction of persistence MAE0.06134196370543393independently computed from committed summaries (method experiments/cp2_power_v1/METHOD.md); inputs reportedexperiments/cp2_power_v1/results/power_metrics.json @ fa88153ed383 · /timesfm/regions/0/sensitivity/mde_fraction_of_mae_persistence/power_0.80
EDL smallest effect size of interest (frozen contract)0.0001independently computed from committed summaries (method experiments/cp2_power_v1/METHOD.md); inputs reportedexperiments/cp2_power_v1/results/power_metrics.json @ fa88153ed383 · /edl/constants/sesoi
EDL alpha (one-sided)0.05independently computed from committed summaries (method experiments/cp2_power_v1/METHOD.md); inputs reportedexperiments/cp2_power_v1/results/power_metrics.json @ fa88153ed383 · /edl/constants/alpha
EDL K01: committed sessions293independently computed from committed summaries (method experiments/cp2_power_v1/METHOD.md); inputs reportedexperiments/cp2_power_v1/results/power_metrics.json @ fa88153ed383 · /edl/contrasts/0/aggregate/committed_sessions
EDL K01: power at the SESOI (one-sided alpha 0.05)0.08793576060301156independently computed from committed summaries (method experiments/cp2_power_v1/METHOD.md); inputs reportedexperiments/cp2_power_v1/results/power_metrics.json @ fa88153ed383 · /edl/contrasts/0/aggregate/alpha_one_sided/power_at_sesoi
EDL K01: minimum detectable effect at 80% power0.000853645832344737independently computed from committed summaries (method experiments/cp2_power_v1/METHOD.md); inputs reportedexperiments/cp2_power_v1/results/power_metrics.json @ fa88153ed383 · /edl/contrasts/0/aggregate/alpha_one_sided/mde/power_0.80
EDL K01: sessions for 80% power at the SESOI (planning magnitude)21352independently computed from committed summaries (method experiments/cp2_power_v1/METHOD.md); inputs reportedexperiments/cp2_power_v1/results/power_metrics.json @ fa88153ed383 · /edl/contrasts/0/aggregate/alpha_one_sided/sessions_needed/power_0.80

TimesFM 2.5 zero-shot versus last-close persistence

negative
  • containment_qualified: false
  • exploratory, descriptive; opened periods; not confirmatory
  • evidence_label: reported

Does one pinned TimesFM 2.5 checkpoint forecast the six-bar-ahead endpoint better than assuming no change (last-close persistence)? The primary measure is the paired MAE difference, model minus persistence, in the holdout and fresh regions, reported separately. A positive value means the model was worse. The bootstrap interval is exploratory, not a confirmatory confidence interval.

TimesFM 2.5 zero-shot versus last-close persistence: values with source and evidence label
ValueResultEvidenceSource
containment_qualifiedfalsereportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /containment_qualified
Source labelexploratory, descriptive; opened periods; not confirmatoryreportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /source_label
Run date2026-11-01reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /run_date
holdout: primary, MAE model minus persistence (above 0 = model worse)9.049291580353995e-05reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/24/value
holdout: bootstrap interval p05 (exploratory)6.812455443949697e-05reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/24/ci/lower_p05
holdout: bootstrap interval p95 (exploratory)0.0001130125802933661reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/24/ci/upper_p95
holdout: sessions185reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/24/n_sessions
holdout: MAE, TimesFM0.0018584222789986195reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/25/value
holdout: MAE, last-close persistence0.0017679293631950796reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/26/value
fresh: primary, MAE model minus persistence (above 0 = model worse)0.00015244070356988247reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/0/value
fresh: bootstrap interval p05 (exploratory)8.82182977763805e-05reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/0/ci/lower_p05
fresh: bootstrap interval p95 (exploratory)0.00021964950229628156reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/0/ci/upper_p95
fresh: sessions15reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/0/n_sessions
fresh: MAE, TimesFM0.0019771235873262297reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/1/value
fresh: MAE, last-close persistence0.0018246828837563472reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/2/value
Verdict (claims ledger) (C04-R17)negativereportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R17/value_as_written

EDL: development-period diagnostics

K01: positive exploratory diagnostic; K02–K10, K16–K22: inconclusive
  • containment_qualified: false
  • EDL E2 development-period metrics written by the cp2_edl_runner_v1 E2 runner on the owner's PC; exploratory, assumption-conditional diagnostics; not confirmatory
  • evidence_label: reported

Exploratory, assumption-conditional diagnostics of the EDL trading-rule candidates on the development period only, in paper-money units (fraction of a notional K). Contrast K01 compares candidate C02 with C01. No holdout or fresh EDL result exists in a committed file. Not confirmatory; no study-wide error rate is certified.

EDL: development-period diagnostics: values with source and evidence label
ValueResultEvidenceSource
containment_qualifiedfalsereportedexperiments/cp2_tfm_edl_results_v1/results/edl/edl_metrics.json @ fa88153ed383 · /containment_qualified
Regiondevelopmentreportedexperiments/cp2_tfm_edl_results_v1/results/edl/edl_metrics.json @ fa88153ed383 · /region
Run date2026-11-01reportedexperiments/cp2_tfm_edl_results_v1/results/edl/edl_metrics.json @ fa88153ed383 · /run_date
K01 (C02−C01): development summary, fraction of K0.0010162989890025384reportedexperiments/cp2_tfm_edl_results_v1/results/edl/edl_metrics.json @ fa88153ed383 · /edl_contrasts/0/development_summary
K01: empirical p (one-sided)0.0015369660828796455reportedexperiments/cp2_tfm_edl_results_v1/results/edl/edl_metrics.json @ fa88153ed383 · /edl_contrasts/0/empirical_p
K01: Holm-adjusted p0.04457201640350972reportedexperiments/cp2_tfm_edl_results_v1/results/edl/edl_metrics.json @ fa88153ed383 · /edl_contrasts/0/adjusted
K01: fold 4 mean-0.0004321240777458181reportedexperiments/cp2_tfm_edl_results_v1/results/edl/edl_metrics.json @ fa88153ed383 · /edl_contrasts/0/folds/4/mean
One-SE selection: statusQUALIFIEDreportedexperiments/cp2_tfm_edl_results_v1/results/edl/edl_metrics.json @ fa88153ed383 · /edl_one_se/status
One-SE selection: selected candidateC02reportedexperiments/cp2_tfm_edl_results_v1/results/edl/edl_metrics.json @ fa88153ed383 · /edl_one_se/selected
K01 verdict (claims ledger) (C04-R41)positive exploratory diagnosticreportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R41/value_as_written
K02–K10, K16–K22 verdict (claims ledger) (C04-R44)inconclusivereportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R44/value_as_written
Holdout and fresh EDL families (claims ledger) (C04-R55)NO QUALIFIED EDL EMPIRICAL RESULTSreportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R55/value_as_written

Regime decoding check

Historical regime labels were decoded with the whole sequence, which quietly uses later bars. The MR-01 study re-decoded every bar causally, using only the past, and compared the two.

Causal and whole-sequence labels agree on 0.9861179970252851 to 0.9985443959243085 of rowslabel agreement, share of rows · gate on the right
0.9850.9900.9951.000QQQM · dev fold 1PASSQQQM · dev fold 2PASSQQQM · dev fold 3PASSQQQM · dev fold 4PASSQQQM · holdoutPASSQQQM · freshPASSXAUUSD · dev fold 1PASSXAUUSD · dev fold 2PASSXAUUSD · dev fold 3PASSXAUUSD · dev fold 4PASSXAUUSD · holdoutPASSXAUUSD · freshPASS 0.991.00QQQM · dev fold 1PASSQQQM · dev fold 2PASSQQQM · dev fold 3PASSQQQM · dev fold 4PASSQQQM · holdoutPASSQQQM · freshPASSXAUUSD · dev fold 1PASSXAUUSD · dev fold 2PASSXAUUSD · dev fold 3PASSXAUUSD · dev fold 4PASSXAUUSD · holdoutPASSXAUUSD · freshPASS
Lowest: XAUUSD historical fresh; highest: QQQM development fold 4. The axis does not start at zero. Change in Macro-F1 between the two labelings: not in the results export. Source: experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383; evidence reported (owner-run; metrics file as posted; not recomputed). sensitivity analysis of a stated limitation; exploratory; not a new model, not confirmatory
Values in this chart, with source and evidence
MR-01 values used in this chart
ValueResultEvidenceSource
QQQM development fold 1: label agreement, causal vs whole-sequence0.9958978430594151reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_1/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
QQQM development fold 1: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_1/status
QQQM development fold 2: label agreement, causal vs whole-sequence0.9943099113404791reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_2/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
QQQM development fold 2: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_2/status
QQQM development fold 3: label agreement, causal vs whole-sequence0.9915310308323408reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_3/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
QQQM development fold 3: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_3/status
QQQM development fold 4: label agreement, causal vs whole-sequence0.9985443959243085reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_4/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
QQQM development fold 4: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1development_fold_4/status
QQQM fixed holdout: label agreement, causal vs whole-sequence0.9939266701657045reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1fixed_holdout/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
QQQM fixed holdout: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1fixed_holdout/status
QQQM historical fresh: label agreement, causal vs whole-sequence0.9919028340080972reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1historical_fresh/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
QQQM historical fresh: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/QQQM~1historical_fresh/status
XAUUSD development fold 1: label agreement, causal vs whole-sequence0.9983453709783322reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_1/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
XAUUSD development fold 1: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_1/status
XAUUSD development fold 2: label agreement, causal vs whole-sequence0.9965331582403152reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_2/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
XAUUSD development fold 2: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_2/status
XAUUSD development fold 3: label agreement, causal vs whole-sequence0.9977675640183847reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_3/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
XAUUSD development fold 3: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_3/status
XAUUSD development fold 4: label agreement, causal vs whole-sequence0.9962967826657912reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_4/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
XAUUSD development fold 4: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1development_fold_4/status
XAUUSD fixed holdout: label agreement, causal vs whole-sequence0.9903559947612811reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1fixed_holdout/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
XAUUSD fixed holdout: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1fixed_holdout/status
XAUUSD historical fresh: label agreement, causal vs whole-sequence0.9861179970252851reported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1historical_fresh/variants/causal_online_viterbi_cold/label_agreement_with_whole_sequence
XAUUSD historical fresh: replication gatePASSreported (owner-run; metrics file as posted; not recomputed)experiments/cp2_mr01_causal_v1/results/mr01_metrics.json @ fa88153ed383 · /populations/XAUUSD~1historical_fresh/status

Adaptive vs global

In the same files, the regime-routed model's Macro-F1 sits beside the single global model's, population by population. Development values are not trading evidence.

Regime routing next to the global model, population by populationMacro-F1 · whole-sequence decoding
0.340.360.380.400.42QQQM · dev meanglobal: not in the results exportQQQM · holdoutQQQM · freshXAUUSD · dev meanglobal: not in the results exportXAUUSD · holdoutXAUUSD · fresh 0.350.40QQQM · dev meanQQQM · holdoutQQQM · freshXAUUSD · dev meanXAUUSD · holdoutXAUUSD · fresh
Values as the claims ledger records them; the global comparator is relisted unchanged from the saved baseline table. Per-fold development values: not in the results export. Global development value: not in the results export. Number of populations in which the adaptive model is ahead: not in the results export. Evidence reported; development values are not trading evidence.
Values in this chart, with source and evidence
Adaptive and global Macro-F1 values used in this chart
ValueResultEvidenceSource
QQQM development (equal mean of the four folds) whole-sequence Macro-F1 (C04-R179)0.3896784170912872reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R179/value_as_written
QQQM fixed holdout whole-sequence Macro-F1 (C04-R99)0.391185042763235reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R99/value_as_written
QQQM fixed holdout global Macro-F1 (relisted) (C04-R105)0.3559988734227451reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R105/value_as_written
QQQM historical fresh whole-sequence Macro-F1 (C04-R109)0.3943384106077888reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R109/value_as_written
QQQM historical fresh global Macro-F1 (relisted) (C04-R115)0.3624472129438476reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R115/value_as_written
XAUUSD development (equal mean of the four folds) whole-sequence Macro-F1 (C04-R185)0.3439477396205161reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R185/value_as_written
XAUUSD fixed holdout whole-sequence Macro-F1 (C04-R159)0.40280872264291845reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R159/value_as_written
XAUUSD fixed holdout global Macro-F1 (relisted) (C04-R165)0.3634154387733373reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R165/value_as_written
XAUUSD historical fresh whole-sequence Macro-F1 (C04-R169)0.37821750685585914reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R169/value_as_written
XAUUSD historical fresh global Macro-F1 (relisted) (C04-R175)0.3465026681071915reportedexperiments/cp2_report_draft2_v1/manuscript_v1/CH04/CH04_RESULTS_CLAIMS.csv @ fa88153ed383 · C04-R175/value_as_written

Foundation-model test

A separate, pre-registered test asked whether one pinned TimesFM 2.5 checkpoint, used zero-shot, forecasts the six-bar-ahead return better than simply assuming no change. The result is published as found.

TimesFM's error next to no-change persistence in both periodsmean absolute error, as written · lower is better
holdoutn 1850.00176792936319507960.0018584222789986195freshn 150.00182468288375634720.0019771235873262297 holdoutn 185freshn 15
holdout: model minus persistence 9.049291580353995e-05 (above 0 = model worse). fresh: model minus persistence 0.00015244070356988247 (above 0 = model worse). MAE is shown in the units of the committed file; the board's basis-point labels are not in the results export. Evidence reported; containment_qualified: false · exploratory, descriptive; opened periods; not confirmatory · evidence_label: reported.
Values in this chart, with source and evidence
TimesFM and persistence values used in this chart
ValueResultEvidenceSource
holdout: MAE, last-close persistence0.0017679293631950796reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/26/value
holdout: MAE, TimesFM0.0018584222789986195reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/25/value
holdout: primary, MAE model minus persistence (above 0 = model worse)9.049291580353995e-05reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/24/value
holdout: sessions185reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/24/n_sessions
fresh: MAE, last-close persistence0.0018246828837563472reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/2/value
fresh: MAE, TimesFM0.0019771235873262297reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/1/value
fresh: primary, MAE model minus persistence (above 0 = model worse)0.00015244070356988247reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/0/value
fresh: sessions15reportedexperiments/cp2_tfm_edl_results_v1/results/timesfm/timesfm_metrics.json @ fa88153ed383 · /metrics/0/n_sessions

Limitations

  • Target-maturity purge not established (MR02).The inspected original adjacent split functions did not establish an explicit target-maturity purge for overlapping targets. No affected-row count is calculated and no performance correction is claimed.

    Source: experiments/cp2_report_draft2_v1/manuscript_v1/CH03/CH03_DRAFT2.md §3.2.2 · source-inspected

  • Exploratory, not confirmatory.No study-wide error rate is certified. The separate EDL study found one positive exploratory diagnostic in development; it is not confirmed.
  • Synthetic checks are only that.Synthetic fits verify the regime method's recovery; they say nothing about real markets.
  • Development numbers are not trading evidence.Costs are fixed assumptions; no live trading took place.
  • Retrospective regime decoding (MR-01).Original whole-window decoding lets later evaluation emissions influence earlier state assignments. The original result is retrospective conditional classification evidence, not established row-by-row causal adaptive superiority.

    Source: experiments/cp2_report_draft2_v1/manuscript_v1/CH03/CH03_DRAFT2.md §3.2.5 · source-inspected

  • Descriptive, not general.The two instruments have different observed histories and sessions. Their comparison is descriptive across these two datasets, not a general experiment over gold and equity asset classes.

    Source: experiments/cp2_report_draft2_v1/manuscript_v1/CH03/CH03_DRAFT2.md §3.1.1 · source-inspected

  • No trading claim.Neither predictive robustness nor decision support implies that the system is profitable or deployment-ready. The decision-support prototype is paper-only and executes no broker orders.

    Source: experiments/cp2_report_draft2_v1/manuscript_v1/CH03/CH03_DRAFT2.md §3.1.1 · source-inspected

Model cards

HMM · XAUUSDFour states, diagonal covariance, frozenHMM · QQQMFour states, diagonal covariance, frozenRegime XGBoost familyOne classifier per regime and instrumentGlobal XGBoost fallbackOne classifier for all bars

References

  • Andrew Ang; Geert Bekaert (2002). International Asset Allocation with Regime Shifts. The Review of Financial Studies, 15(4), 1137–1187. DOI 10.1093/rfs/15.4.1137.

  • Matthew Wang; Yi-Hong Lin; Ilya Mikhelson (2020). Regime-Switching Factor Investing with Hidden Markov Models. Journal of Risk and Financial Management, 13(12), 311. DOI 10.3390/jrfm13120311.

  • Abhimanyu Das; Weihao Kong; Rajat Sen; Yichen Zhou (2024). A decoder-only foundation model for time-series forecasting. ICML; Proceedings of Machine Learning Research, 235, 10148–10167.

  • Christoph Bergmeir; Rob J. Hyndman; Bonsoo Koo (2018). A note on the validity of cross-validation for evaluating autoregressive time series prediction. Computational Statistics & Data Analysis, 120, 70–83. DOI 10.1016/j.csda.2017.11.003.

  • Vitor Cerqueira; Luis Torgo; Igor Mozetič (2020). Evaluating time series forecasting models: an empirical study on performance estimation methods. Machine Learning, 109, 1997–2028. DOI 10.1007/s10994-020-05910-7.

  • Sokolova, Marina; Lapalme, Guy (2009). A systematic analysis of performance measures for classification tasks. Information Processing & Management45(4),427–437. DOI10.1016/j.ipm.2009.03.002.

A selection from the citation ledgers of the CP2 report draft. The full list with links is on the References and data sources page. Every number on this page is read from a committed result file at build time; none is typed by hand.