The Impossible Trinity of Time-Series Validation: A Conservation Law among Training Sufficiency, Test Coverage, and Temporal Causality

1.1 The problem

• The paper formalizes training sufficiency, test coverage, and temporal causality as competing requirements for time-series validation.Sufficiency concerns training-set size in each run, coverage concerns the union of test sets, and causality requires training data to precede test data.Source: §1.1 · pp. 1–2

1.2 Contributions

• Every scheme obeys a ledger linking sufficiency, coverage, and future training mass, while a separate proximity constraint and mixing-based bound price leakage severity.Strict causality requires α + β ≤ 1; the contributions also state that leakage bias under β-mixing is bounded by 2Mβmix(δ).Source: §1.2 · p. 2
• The paper assigns costs to each validation trade-off: learning-curve bias for insufficient training, variance and regime error for limited coverage, and memory-dependent costs for causality.An embargo can buy back causality under finite memory at sample cost O(m(2H + h)/T), while the non-stationary component is not redeemable.Source: §1.2 · p. 2
• The paper places standard schemes in shared coordinates and includes a numerical illustration of the difference between anti-causal mass and effective leakage.Non-trivial proofs are collected in Appendix A.Source: §1.2 · p. 3

2.1 Data and validation schemes

• A validation scheme consists of training/test folds on time-indexed feature-label samples, and its reported estimate is a weighted average of fold test losses.Labels may resolve H periods after their sample index, so index order need not equal information order.Source: §2.1 · p. 3, §2.2 · p. 3

2.2 The three quantities

• Training sufficiency is measured by worst-fold and average training fractions, with the worst fold used to represent the scheme’s weakest training run.The average uses the weights defining the reported estimate.Source: §2.2 · p. 3
• Strict causality requires every fold’s training set to precede its test set and is equivalent to having no future training data at any test point.Under this condition, the trained model is measurable using information available before testing, so the fold is an adapted predictor.Source: §2.2 · p. 4
• Anti-causal mass measures how much future training data is used, while the anti-causal margin measures how close that data is to a test point.The margin is the distance to the first future training sample encountered from a test point.Source: §2.2 · p. 4

2.3 Label horizon and informational order

• When labels resolve after a horizon H, training must leave a gap so it uses only resolved labels and avoids overlapping test-label windows.Purging removes training samples whose label windows overlap test-label windows; the combinatorial statements receive an O(mH/T) correction when H > 0.Source: §2.3 · p. 4

2.4 The dependence structure of the process

• The statistical theorem assumes uniformly bounded β-mixing dependence, but does not require stationarity for its main dependence bound.Stationarity enters later when fold estimates are identified with a learning curve.Source: §2.4 · p. 5
• The framework compares the observed loss with leak-free references that independently replace future training data or remove leakage on both sides.The fully decoupled reference replaces past training data, the test point, and future training data with mutually independent copies having the same marginal laws.Source: §2.4 · p. 5

3 The main theorem and its combinatorial consequences

• The theorem’s ledger, proximity bound, and trinity inequality jointly constrain sufficiency, coverage, and the distance of future training data.Any excess of α + β over 1 requires anti-causal mass, and when future data are used its distance from a test point is at most (1 − α)T.Source: §3 · pp. 5–6
• Under β-mixing, leakage harm depends on distance to future training data rather than its volume; purging and an embargo trade that distance against sufficiency.Distance is cheap when the process forgets quickly.Source: §3 · p. 6

3.1 Combinatorial consequences

• Expanding walk-forward attains every point on the strictly causal Pareto frontier, and every strictly causal scheme is weakly dominated by one of these schemes.The frontier is α + β = 1.Source: §3.1 · pp. 6–7
• With full test coverage, a strictly causal scheme’s average training fraction is at most about one half, and the first test block of the expanding scheme has no training data.At β = 1, the worst-fold training fraction is α = 0.Source: §3.1 · p. 8
• Contiguous-block k-fold is an extremal buyer of future data: it uses the minimum anti-causal mass required by its sufficiency and coverage coordinates, all at margin δ = 1.Its margin cannot be pushed beyond T/k without giving up sufficiency.Source: §3.1 · p. 8
• A label horizon or embargo consumes training sufficiency: with horizon H, the causal bound tightens by H/T, and finer blocking increases sanitization overhead.The stated total correction for m train/test boundaries is O(mH/T).Source: §3.1 · p. 8

4.1 The price of causality: the severity law

• Leakage bias depends on the nearest future-training margin δ+, not just anti-causal mass Λ; placing the same Λ beyond h caps harm at 2Mβmix(h).The total-variation bound also admits a pathwise coupling interpretation: future training data can be replaced by an independent copy with probability at least 1 − βmix(δ+).Source: §4.1 · pp. 8–9

4.2 The price of sufficiency: learning-curve bias

• Under Assumption B, smaller training folds yield pessimistic learning-curve bias; comparisons at αT ≪ T systematically favor simpler models when learning curves have different slopes.This follows because complex models’ learning curves can decay more slowly; monotonicity of the learning curve is a genuine restriction, though the proposition can use range extrema on [αT, T].Source: §4.2 · pp. 10–11

4.3 The price of coverage: variance and regime sampling

• The coverage gap costs variance of order σ2/(βT) and can cause irreparable worst-case regime error when risk differs on intervals absent from the observed loss record.Two processes can agree on the covered losses yet differ on an uncovered period, leaving any estimator based on those losses unable to recover the full time-averaged risk.Source: §4.3 · p. 11

4.4 A statistical impossibility theorem

• Strictly causal schemes cannot eliminate both learning-curve bias and coverage error, while escaping that floor with a non-causal scheme requires future training mass that incurs leakage bias.The three estimation-error components are leakage bias, learning-curve bias, and variance plus regime error.Source: §4.4 · p. 12

4.5 The hardness of the trinity is the memory of the process

• In the i.i.d. limit with H = 0, βmix(d) = 0 for d ≥ 1, so anti-causal training does no statistical harm; autoregressive validity can instead reflect an algorithm-specific small exchange rate.The trinity arises from dependence and non-stationarity, not from the time-series label alone.Source: §4.5 · p. 12
• With exponential mixing, purged k-fold plus an embargo trades sanitization distance h for leakage at most 2MCe^-h/τ; the window closes when τ or H approaches T/k.This repairs the stationary-dependence part of causality, but not drift: a future-trained model may see the regime it is being tested on, unlike the deployed model.Source: §4.5 · pp. 12–13

5.1 Coordinates

• Expanding and rolling walk-forward match in (α, β), but their average training fractions differ by up to (1 − w)/2.Rolling windows trade learning-curve bias for robustness to drift, which the stationary framework cannot price.Source: §5.1 · p. 13
• CPCV produces many backtest paths while purging and embargo control Λeff, at the cost of reduced sufficiency and more computation.This changes coverage's meaning by sampling regimes and paths, rather than only test points.Source: §5.1 · p. 13

5.2 The multiplier effect of tuning

• Selecting among N configurations amplifies variance into optimism; low coverage raises variance and contributes to backtest-overfitting risk.Selection can also favor configurations that exploit leakage and distort rankings when learning-curve slopes differ.Source: §5.2 · p. 14
• Tuning should be nested: select parameters inside, then confirm on an outer strictly causal segment untouched by selection.Reported metrics should also be corrected for multiplicity in N.Source: §5.2 · p. 14

5.3 A numerical illustration of volume versus severity

• The numerical illustration uses i.i.d. returns with pure-noise labels and 1-NN regression on past-only, temporally smooth features.The walk-forward, hold-out, and purged 5-fold schemes use an H-gap; shuffled and contiguous 5-fold share (α, β, Λ) = (0.8, 1, 0.8), up to O(1/T) for shuffled Λ.Source: §5.3 · pp. 14–15
• Spurious skill decays monotonically with exclusion radius g and reaches statistical zero before g ≈ H.Figure 2 plots the 1-NN information coefficient against g alongside the label-dependence envelope.Source: §5.3 · p. 14, Figure 2 · p. 15
• IC was +0.318 ± 0.012 for shuffled 5-fold, versus +0.004 ± 0.012 for contiguous 5-fold, despite identical (α, β, Λ) up to O(1/T).The schemes differ in margin placement: shuffled folds put anti-causal mass beside every test point, while contiguous folds do so only near O(k) boundaries.Source: §5.1 · p. 13, §5.3 · p. 15

• Estimate the label horizon H and dependence scale τ first; express gaps, purges, and embargoes in H + cτ, with c ≈ 2–3.The suggested dependence ruler is the autocorrelation half-life; financial data may also have volatility clustering.Source: §6 · p. 15
• When stationarity is credible and the goal is comparison or tuning, use purged k-fold with embargo or CPCV; when non-stationarity is feared, use expanding walk-forward.For walk-forward, report per-fold performance against training size and rehearse tuning inside the rolling procedure.Source: §6 · p. 16
• Nest tuning: make inner selection separate from an outer strictly causal confirmation segment, and correct the selected winner for multiplicity in N.The guidance also names the deflated Sharpe ratio and PBO as options.Source: §6 · p. 16
• Report (α, ᾱ, β, Λ, δ) and avoid shuffled k-fold on data with serial dependence or overlapping labels.The proposed red line identifies small-δ leakage, rather than alpha, as the first suspect in a too-good backtest.Source: §6 · p. 16

• The h-block and hv-block methods formalized deleting neighborhoods around test points; the paper's exclusion-radius sweep is hv-block with a varying radius.Burman et al. (1994); Racine (2000).Source: §7 · p. 16
• Prior work established k-fold validity for specific autoregressive settings and compared standard schemes empirically; purging, embargo, and CPCV were systematized by Lόpez de Prado (2018).The paper casts embargo as a theorem: its width follows the mixing-coefficient decay scale, and its sufficiency cost is stated separately.Source: §7 · p. 16
• The stated contribution is a proof that all schemes obey one inequality, with a Pareto frontier and a price list for violations.This is framed as distinct from proposing another validation scheme.Source: §7 · p. 16
• Generalization bounds for non-stationary mixing processes address a different object than this paper's validation-scheme design space; the two are complementary.The discussion also connects low-coverage selection inflation to the Reality Check of White (2000), the deflated Sharpe ratio, and PBO.Source: §7 · pp. 16–17

• Every scheme obeys α + min{β, δ/T} ≤ 1, independent of all statistical assumptions; on its causal face α + β ≤ 1, the walk-forward family is the exact Pareto frontier.The only exit is anti-causal training mass, with Λ ≥ α + β − 1, placed within (1 − α)T of the tests.Source: §8 · p. 17
• Expanding walk-forward is the exact Pareto frontier of strictly causal validation, while escaping that frontier requires future training data.The amount of anti-causal mass must cover the excess of α + β over 1.Source: §8 · p. 17
• For graph-dependent data, the margin becomes graph distance, and graph-neighborhood growth sets the distance-to-sufficiency exchange rate.The conclusion gives examples ranging from order δ on a line to exponential growth in δ on an expander.Source: §8 · p. 17

• Proofs of Corollaries 2 and 3, Proposition 6, and Proposition 7(i) are indicated in the text and omitted as immediate.Source: A Proofs · p. 17

• βT ≤ T − p + 1, with equality exactly when U = [p, T].Source: §A.1 · p. 17
• The proof characterizes equality in α + β ≤ 1 + Λ through the fold’s leakage, training-set size and contents, and the test-set endpoint.For a fold whose earliest test point is p, Λ_iT counts its training points strictly after p.Source: §A.1 · pp. 17–18
• When future training is present, α + δ/T ≤ 1; strictly causal schemes satisfy α + β ≤ 1, attained on the frontier by expanding walk-forward.Source: §A.1 · p. 18

• For bounded loss |ψ| ≤ M, the proof compares the joint law of past, test, and future blocks with independent blocks using total variation.Source: §A.2 · p. 18
• Berbee coupling replaces the future block with an independent copy, with mismatch probability bounded by βmix(δ+).Source: §A.2 · p. 18
• The pointwise leakage bound is 2M[βmix(δ−) + βmix(δ+)]; averaging over test points gives the fold-level inequalities.Source: §A.2 · p. 18

A.3 Theorem 4 (mean sufficiency)

• For disjoint equal-length test intervals, their positions constrain the training-set sizes; the equal-block expanding walk-forward construction makes each inequality tight.Source: §A.3 · pp. 18–19

A.4 Proposition 7(ii) (regime term)

• When two risk profiles differ by Δ outside U and agree on U, their targets differ by (1 − β)Δ although any estimator has the same distribution under both processes.Source: §A.4 · p. 19

A.5 Corollary 8 (statistical impossibility)

• The corollary gives a positive risk floor at β = 1 and adds a non-causal term using the leakage-bias bound.The floor uses strict decrease of L on [T/2, T]; the non-causal term follows by combining Theorem 1(a) and Theorem 1(d).Source: §A.5 · p. 19