acceptodds
Under review as a conference paper at ICLR 2027

WHEN THERE IS NOTHING TO PREDICT: DIRECTIONAL BIAS IN TIME SERIES FOUNDATION MODELS

Abstract

Time series foundation models forecast a new series from its context alone. We ask what they forecast when there is nothing to predict. On a martingale the last observed value is the best point forecast under squared loss, and any other forecast adds risk equal to the mean square of its departure from that value, so what a model adds can be measured against a known optimum. The part of a departure that stays the same when the history is negated cannot come from the direction of the history; removing it is the mirror correction. We find that several released models carry a directional bias that, on rising markets, looks like skill. On random walks five of the eleven models, none of them decoder-only, forecast a rise on up to 85 percent of contexts. A corpus with positive trends suffices to cause such a bias, in a decoder-only model as well. On daily prices every model loses to the last value once the return signs are randomised, and on equity windows from a rising market the mirror correction lowers 128-day skill for ten of the eleven models. On 1996–2014 windows several models time the market far better than trend rules read from the same contexts, which points to memory: public training data contain those years. The mirror correction, already built into TimesFM-2.5, lowers risk whenever past and future are symmetric under negation, and an exact identity gives its cost elsewhere; mirror augmentation during fine-tuning removes most of Chronos-Bolt's bias.

Primary area:
learning on time series and dynamical systems
Keywords:
time series analyse, foundation model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_EsHtOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates directional bias in time series foundation models. Through synthetic and financial experiments, it shows that upward bias can resemble forecasting skill in rising markets, examines possible historical memorization, and evaluates mirror correction and augmentation.

Critical strengths

The paper identifies an interesting and relevant phenomenon.

The experimental design is clever, with carefully chosen controls that help distinguish several confounding factors.

Critical weaknesses

The theoretical arguments are largely standard and lack clear novelty.

The findings are not sufficiently translated into concrete guidance for TSFM pretraining and downstream applications.

Multivariate forecasting and the use of covariates receive insufficient discussion.

Questions for authors

Please develop concrete implications for TSFM pretraining or downstream applications, supported by additional analysis or experiments.

Please investigate the role of covariates and discuss whether the findings extend to multivariate forecasting.

Please position the theoretical results as supporting analysis and moderate the emphasis on theoretical novelty.

I would be willing to increase my score if the authors satisfactorily address these three points.

5 comments
AuthorsOpenReview ↗

Response to Reviewer EsHt (1 of 2): Q3, W1, Q1 and W2, theory and practical guidance

Thank you for the review and for stating clearly what would change your assessment.

To Q3 and W1, Thank you for asking us to position the theory as supporting analysis. We agree that the arguments are standard, and the revision presents them that way.

  • The abstract no longer mentions the theory.
  • The introduction says that both propositions rest on standard arguments and serve as measuring tools, and it cites earlier work for each one.
  • Section 2.1 calls Proposition 1 the standard decomposition of squared error about a conditional mean. It adds that we claim no novelty for it and use it as a measuring device.
  • Section 2.4 says that parts (i) and (ii) of Proposition 2 restate the gain from averaging over a symmetry group, and that part (iii) is an exact accounting identity used to read the experiments.

To Q1 and W2, Thank you for asking for concrete guidance supported by additional analysis or experiments. The paper closes with guidance for downstream use and for pretraining (Section 7, Appendix I). For downstream use the revision adds two checks, a backtest on years after the training data and a test with an unrelated covariate. New experiments support all four checks. The guidance for pretraining rests on the corpus and fine-tuning experiments from the submission (Figure 4, Tables 15 and 30).

For downstream use, Section 7 lists checks that need only forecasts.

  • Compare the model with trivial forecasts, namely the last value, a drift fitted to the history and, for direction, a forecast that always rises. In the new backtest no model's direction strategy beats holding the market by two standard errors (Table 2). On rising equities no model calls the direction more often than a forecast that always rises (Table 23).
  • Score the model on sign-randomised copies of the same windows, in which every step keeps its size and gets a random sign. A copy has no drift by construction, so a systematic departure on it is bias. No model beats the last value on copies of financial windows, nor, in new runs, on copies of electricity, traffic, weather and ETTh1 windows (Appendix N).
  • Backtest on years after the training data. In the new backtest, four models whose forecasts of the 1996-2014 market point to memory have direction Sharpe ratios of 1.46 to 2.31 on those years and 0.00 to 0.70 on 2017-2026 (Table 2). On earlier years, timing that neither reverses on mirrored histories nor survives on copies, or that follows the calendar when the history is shifted, points to memory.
  • Before adding covariates, check that an unrelated rising or falling covariate leaves the direction of the forecasts unchanged. In new runs such a covariate changes the forecasts of Chronos-2 and Moirai-2.0 (see our next comment).

The checks complement each other. In squared error alone, the first check would not flag Chronos-T5, Chronos-Bolt and TiRex, whose point estimates are above both the last value and the fitted drift on raw daily windows at 128 steps (Table 3). Its direction part does flag them, and so does the second check, on which all three depart upward.

For pretraining, Section 7 makes these recommendations.

  • Report the trend statistics of the corpus. A corpus with positive trends was enough to give both of our small models an upward bias (Figure 4). In three public corpora 52 to 57 percent of 128-step moves are rises, and almost all of the excess comes from the M4 subsets (Table 15). These statistics show where a bias can come from. Whether the corpus is the cause in a released model depends on how the model weights its subsets, which is not public (Section 4).
  • Mirror augmentation during fine-tuning removed most of Chronos-Bolt's bias. It brought Chronos-Bolt's share of rising forecasts on random walks to 0.58, against 0.71 without it and 0.79 as released, with no loss on ETTh1. It also lowers skill on rising equities, to 0.064 against 0.126 (Table 30).
AuthorsOpenReview ↗

Response to Reviewer EsHt (2 of 2): Q2 and W3, covariates and multivariate forecasting

To Q2 and W3, Thank you for asking about covariates and multivariate forecasting. The submission used neither, and the revision adds a study of both (Section 4, Appendix M). A covariate can bring a direction of its own, and more channels do not remove the bias in general.

(a) Four of the eleven models accept further inputs in the interfaces we use. Chronos-2 and Moirai-1.1 take joint targets, past covariates and known future covariates. Moirai-2.0 takes past and known future covariates. TimesFM-2.5 takes known future covariates through a linear regression.

(b) In two of the four models, an unrelated covariate changes the direction of the forecast. We paired each random walk with a covariate that is independent of it, so the last value remains the best forecast, and let the covariate drift up or down by 0.1 standard deviations per step. Replacing a falling covariate by a rising one raises the share of rising forecasts at 128 steps by 0.40 for Chronos-2 and 0.54 for Moirai-2.0 (Table E1). Moirai-1.1 barely changes. TimesFM-2.5, given an independent or a rising known future covariate, stays at 0.50 to 0.52.

(c) Two positive controls check whether a model uses the channel at all. With a known future covariate that drives the target, Chronos-2 reaches the attainable skill of 0.64 and TimesFM-2.5 about half of it. Moirai-2.0 stays behind the last value there (−0.11), although an unrelated covariate moved its forecasts in (b). No model uses a past covariate that leads the target by one step, and with it Moirai-1.1 is far behind the last value (skill −1.90).

(d) With several inputs, what the mirror correction removes depends on which inputs are mirrored. If the target and the covariate are negated together, the shift caused by the covariate changes sign with them and stays in the odd part, which the correction keeps. If the target alone is negated, the shift is the same in both forecasts and falls into the even part, which the correction removes (Table 37). The mirror therefore has to be applied to the inputs whose direction is in question.

(e) For multivariate forecasting, Chronos-2's share of rising forecasts falls from 0.80 with one channel to about one half with five or more independent channels, and it falls less when the channels are correlated. Moirai-1.1's stays between 0.62 and 0.74 (Table E2). This analysis is post hoc and covers the two models that forecast several targets jointly.

(f) On the financial task we gave the models the market as a past covariate or forecast up to 50 series jointly. With none of these inputs does a direction strategy beat holding the market by two standard errors, and none gives a RankIC above zero after a Holm correction (Table 39). With 50 joint series Moirai-1.1's RankIC falls by 0.056 ± 0.022, and Chronos-2's share of rising forecasts falls from 0.79 to 0.43.

(g) With the real covariates of ETTh1 and Weather, the effect depends on the model and the data (Table 40).

Table E1 (new runs, Table 36 of the revision). Random-walk targets, 128 series. Columns 2 and 3 give the share of rising forecasts at 128 steps without a covariate and with an independent past covariate. Column 4 gives the change in that share when a falling past covariate is replaced by a rising one. Columns 5 and 6 are the positive controls, the skill at 16 steps with a known future covariate that drives the target and the skill at 1 step with a covariate that leads it. Both have an attainable skill of 0.64.

ModelNo covariateIndependent covariateRising minus fallingKnown futureLeading
Chronos-20.700.66+0.40+0.65 ± 0.04−0.06 ± 0.05
Moirai-1.10.690.70−0.02+0.10 ± 0.05−1.90 ± 0.67
Moirai-2.00.590.52+0.54−0.11 ± 0.06−0.02 ± 0.07
TimesFM-2.50.52--+0.34 ± 0.04-

Table E2 (new runs, Table 38 of the revision, post hoc). Share of rising forecasts at 128 steps when several random walks are forecast jointly, by number of channels. Each experiment draws its own 128 random walks, so the one-channel shares differ from Table E1 by sampling.

ModelChannels125102050
Chronos-2independent0.800.580.490.500.510.51
Chronos-2correlated 0.70.710.710.600.600.590.55
Moirai-1.1independent0.650.680.650.640.660.62
Moirai-1.1correlated 0.70.710.670.720.690.740.73

We welcome further discussion if any concerns remain.

Dear Reviewer EsHt,

We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining issues.

Reviewer_u2x9OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper investigates directional bias in time-series foundation models when the underlying series contains no predictable mean structure. Using martingale processes with a known optimal forecast, the authors show that several released TSFMs systematically deviate from the last observed value and exhibit an upward bias. They further introduce a mirror-based decomposition to separate direction-dependent forecasts from direction-invariant bias, and show that positive trends in the pretraining corpus can induce such behavior. Experiments on synthetic processes and financial data suggest that apparent forecasting skill in rising markets can be partly explained by this bias or by memorization of historical patterns.

Critical strengths
  1. Clear diagnostic setting with a known optimum. Martingale processes provide a clean way to distinguish genuine forecasting skill from systematic directional bias.

  2. Strong and interpretable empirical evidence. The study evaluates 11 released TSFMs across synthetic nulls and real financial series, revealing consistent upward bias in several models.

  3. Useful mirror-based analysis and correction. The proposed decomposition separates direction-dependent behavior from direction-invariant bias and is supported by both theory and experiments.

Critical weaknesses

Practical significance beyond diagnosis remains somewhat limited. While the auditing framework provides a useful way to identify directional bias, the proposed correction is not universally safe and the training-time remedy is only validated in limited settings. As a result, the paper offers a strong diagnostic framework but a less complete solution for preventing or correcting such bias in general TSFM deployment.

Calendar-shift evidence appears to be model-dependent. This diagnostic is particularly clear for TimesFM-2.0, whereas the evidence for other models is less conclusive. More systematic validation across different model families would strengthen the memorization argument.

Direction-invariant components should not necessarily be interpreted as harmful bias. When the underlying process contains genuine drift, the even component may also encode useful predictive information. This makes the interpretation of the mirror decomposition less straightforward outside the null setting.

Mirror correction is theoretically guaranteed to reduce risk only under symmetry assumptions. For drifting or asymmetric series, useful predictive information may also be removed, potentially leading to degraded forecasting performance. This limits its applicability as a general-purpose correction strategy.

Empirical evaluation is still largely centered on financial and martingale-like settings. Most conclusions are derived from random walks, exchange rates, and equity returns, with only limited evidence from broader non-financial domains. It therefore remains unclear whether similar directional biases are prevalent in areas such as weather, energy, traffic, or healthcare forecasting.

Questions for authors

Could you provide additional analysis on non-financial real-world time-series domains, such as weather, energy, traffic, or healthcare?

Could the authors further clarify how directional bias should be distinguished from legitimate low-frequency drift in real-world non-martingale series? In such settings, the even component identified by the mirror decomposition may contain both harmful bias and useful predictive signal, making the interpretation of the correction less straightforward.

5 comments
AuthorsOpenReview ↗

Response to Reviewer u2x9 (1 of 3): W1, W3 and W4, what the correction is for

To W1, Thank you for pointing out that the paper offers a stronger diagnosis than remedy. We agree, and the revision now says so.

(a) The mirror correction is presented as a diagnostic (Sections 1 and 6). In the new backtests it does not raise returns (Table 2).

(b) The practical output is the list of four checks that ends Section 7. Each needs only forecasts.

(c) For prevention, mirror augmentation during fine-tuning brings Chronos-Bolt's share of rising forecasts on random walks from 0.71 to 0.58 and keeps its skill on ETTh1 (Table U1). Skill on rising equities falls from 0.126 to 0.064, close to the 0.060 that the mirror correction leaves. The loss on sign-randomised copies is unchanged, so the departures become balanced in sign without becoming smaller. On the positive control, a process on which departing from the last value pays, the model captures 0.540 of the optimal skill against 0.572.

Table U1 (Table 30 of the revision, already in the submission). Chronos-Bolt-small as released, fine-tuned on public training windows, and fine-tuned on the same windows with each one replaced by its mirror with probability one half. Mean ± standard deviation over three seeds.

ProbeReleasedFine-tunedMirror-augmented
Share of rising forecasts, random walks, 128 steps0.790.71 ± 0.030.58 ± 0.04
Skill on ETTh1, 16 steps+0.697+0.697 ± 0.002+0.708 ± 0.003
Skill on raw equities, 128 steps+0.129+0.126 ± 0.018+0.064 ± 0.020
Skill on sign-randomised copies, 128 steps−0.083−0.073 ± 0.006−0.074 ± 0.020
Share of the optimal skill captured on the positive control0.5380.572 ± 0.0070.540 ± 0.005

To W3 and W4, Thank you for noting that the even part can carry useful drift and that the guarantee needs symmetry. We agree with both points, and the paper's numbers show them.

(a) On raw equities at 128 steps, removing the even part lowers skill for ten of the eleven models (all but Moirai-2.0), for four of them by more than two standard errors. For ten models (all but Chronos-2) the even part is an upward drift of 0.4 to 8.7 percent a year (Table 28), on equities that rose 13.6 percent a year. Section 5 therefore reads these gains as the market's drift. They are real gains, and they do not depend on the direction of the history.

(b) The introduction now says that skill removed by the correction may still be useful where the data drift. Section 2.4 gives a known positive drift as an example of an even part that is useful. Appendix I describes the even-energy ratio only as a measure of how symmetric a forecast is.

(c) Sections 1 and 6 state that the correction is guaranteed to lower risk only when past and future are symmetric under negation. Outside symmetry, Proposition 2(iii) gives the exact change in risk as the sum of three terms, the removed energy, the alignment of the even part with the true mean move, and a cross term. For Chronos-T5 on raw equities these change skill by +0.040, −0.102 and −0.004, so the loss comes from the alignment term.

AuthorsOpenReview ↗

Response to Reviewer u2x9 (2 of 3): W2, the memory evidence

To W2, Thank you for noting that the calendar-shift evidence depends on the model. You are right. The check is clear only for TimesFM-2.0. Section 5 now says so, as the limitations already did. For the other three models the mirror and copy checks carry the evidence. Table U2 puts the checks side by side for the four models concerned, Chronos-Bolt, TiRex, Moirai-2.0 and TimesFM-2.0.

(a) On 1996-2014 the four models' mean forecasts follow the realised 128-day moves across 37 end dates, with correlations of 0.71 to 0.87. Drift and momentum rules read from the same histories reach at most 0.26.

(b) Timing that used the direction of the history would reverse on its mirror. Timing that used its magnitudes would survive sign randomisation. For Chronos-Bolt, TiRex and Moirai-2.0 it does neither. On the mirrors it vanishes, and on the copies it falls to about what the best rule reaches. TimesFM-2.0's timing does reverse on the mirror, so this check cannot separate it from a model that reads the history, and the calendar shift matters most for it.

(c) When the end of each history is moved by a few weeks, TimesFM-2.0's forecasts move with the realised futures (0.58, p<0.001). Chronos-Bolt points the same way (0.41). Moirai-2.0 and TiRex (0.33 and 0.29) cannot be separated from a rule that opposes the long-run trend of the history, which reaches 0.29. The p-values of these three models lie between 0.04 and 0.11.

(d) In the new backtests the four models' direction strategies have Sharpe ratios of 1.46 to 2.31 on 1996-2014, against 0.91 or less for the other models. On 2017-2026 they have 0.00 to 0.70, or 0.35 to 0.62 averaged over three rebalancing schedules (a post hoc analysis).

(e) M4 daily, part of several public training corpora, holds series that match the US market's daily returns on 82 percent of the 1996-2014 trading days and on none after October 2017.

(f) Outside finance, a post hoc check adds to the memory evidence for TimesFM-2.0. Its skill of 0.99 at 128 steps on the public Jena weather series falls to −0.02 when each window is reversed in time, while the four other models in the check change by at most 0.06 (Appendix N).

These checks fit memory without proving it, and only the training data could confirm it.

Table U2 (from Tables 2, 21 and 22 of the revision). The four models that time the 1996-2014 market. Columns 2 to 4 give the correlation across end dates between mean forecast and realised 128-day return, from the raw histories, their mirrors and their sign-randomised copies. Column 5 is the calendar-shift correlation, with an asterisk for p<0.01. The last two columns are direction Sharpe ratios.

ModelTiming, rawMirrorCopiesCalendar shiftSharpe 1996-2014Sharpe 2017-2026
Chronos-Bolt+0.78−0.02+0.29+0.41+1.59+0.36
TiRex+0.71−0.17+0.31+0.29+1.78+0.70
Moirai-2.0+0.76−0.05+0.25+0.33+1.46+0.00
TimesFM-2.0+0.87−0.82−0.09+0.58*+2.31+0.34
Best rule read from the history0.26--0.29--
Holding the market----+0.51+0.66
AuthorsOpenReview ↗

Response to Reviewer u2x9 (3 of 3): W5, Q1 and Q2, non-financial domains and drift

To W5 and Q1, Thank you for asking whether the bias appears in non-financial domains. It appears on sign-randomised copies of such series, where the direction is random by construction. We ran the eleven models on electricity, traffic and weather series, ETTh1 and weekly influenza-like-illness (ILI) data (Section 5, Appendix N).

(a) On the copies, each of the five upward-biased models forecasts a rise on 55 to 73 percent of windows at 128 steps, above one half by two standard errors on all four large datasets (Table U3). No model beats the last value there. Moirai-2.0 and FinCast lean down, so a model's own direction need not be upward.

(b) On the raw series the ten general models beat the last value at 16 steps on electricity, traffic and ETTh1 (skill 0.67 to 0.95). The correction changes that skill by at most 0.02 for all ten on electricity and for nine of ten on traffic. These series have strong structure, and the bias is small beside it. Weather is harder. At 16 steps no model exceeds 0.37 there except TimesFM-2.0, which we discuss under W2.

(c) For healthcare we used the weekly ILI series, which gives 77 windows, too few to count as evidence. On its copies only Moirai-1.1 and Moirai-2.0 show a direction.

Table U3 (new runs, Table 41 of the revision). Share of rising forecasts on sign-randomised copies at 128 steps (64 for ILI). An asterisk marks a share that differs from one half by two standard errors. The first five rows are the upward-biased models.

ModelElectricityTrafficWeatherETTh1ILI
Chronos-T50.62*0.59*0.55*0.60*0.51
Chronos-Bolt0.73*0.63*0.62*0.70*0.57
Chronos-20.62*0.55*0.57*0.60*0.46
TiRex0.64*0.56*0.64*0.67*0.52
Moirai-1.10.69*0.69*0.69*0.69*0.65*
Moirai-2.00.480.40*0.44*0.46*0.33*
TimesFM-2.00.48*0.47*0.480.500.49
TimesFM-2.50.480.490.510.500.48
Time-MoE0.500.510.500.510.49
Sundial0.500.510.510.52*0.54
FinCast0.42*0.41*0.41*0.45*0.47

To Q2, Thank you for asking how to tell directional bias from legitimate low-frequency drift on real series. You are right that the even part can hold both, so the mirror decomposition cannot separate them by itself. The revision uses sign-randomised copies for this (Section 5, Appendix N).

(a) A sign-randomised copy keeps the size of every step of a window and draws each sign at random. It has no drift and no seasonality by construction. A systematic departure on the copy therefore cannot have been learned from the series, and we count it as bias.

(b) Only what a model adds on the raw window beyond its departure on the copy can come from the series, drift included.

Table U4 applies the rule to the five upward-biased models on the 900 equity windows. For TiRex and Moirai-1.1 most of the upward departure on real prices is already there on the copies. For Chronos-T5 it is about half and for Chronos-Bolt about a third. For Chronos-2 it is almost none, so by the rule its upward forecasts on these windows can come from the series.

Where a series truly drifts, the even part is a mix, and removing it can cost skill. Proposition 2(iii) measures that cost, and the copies decide what is bias.

Table U4 (from Table 16 of the revision, already in the submission). Mean departure at 128 steps on the 900 equity windows of 2017-2026, on their sign-randomised copies and on the raw windows, in units of the daily return standard deviation of the history, ± standard error.

ModelSign-randomised copiesRaw windows
Chronos-T5+1.46 ± 0.08+2.65 ± 0.25
Chronos-Bolt+1.37 ± 0.13+3.72 ± 0.42
Chronos-2+0.06 ± 0.08+1.93 ± 0.30
TiRex+1.79 ± 0.11+2.19 ± 0.40
Moirai-1.1+2.32 ± 0.18+2.64 ± 0.32

We welcome further discussion if any concerns remain.

Reviewer_9Ts7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates whether existing time-series foundation models contain systematic directional bias. The paper found some models tend to predict under synthetic time-series data without predictable signals. The paper then investigates possible sources and reasons for these biases, and proposes a mirror correction to reduce them. However, the proposed correction only works under specific assumptions and often hurts forecasting performance on real financial data.

Critical strengths

The paper evaluates a wide range of time-series foundation models and conducts several experiments including synthetic data and real-world data to investigate the identified directional bias.

Critical weaknesses

1. Terrible Presentation and Writing

This paper is very hard to read and follow. Many sentences contain too much unclear information as it uses many strange and dense syntax, which makes the sentence very hard to understand even after reading them many times. It also uses many nonstandard invented jargon without sufficient definition and not explaining it when it first shows.

For example, the abstract already contains sentences such as "The part of a departure that stays the same when the history is negated cannot come from the direction of the history; removing it is the mirror correction.", which requires knowing the method in the first place; Otherwise, these sentences are very difficult to understand. Another example will be the word “skill”, which is a very important keyword and already repeatedly been used in the abstract and introduction, but is only introduced and defined formally later in Section 2.3.

The main text of the paper is also overly condensed, often combining methodological changes, numerical results, and statistical conclusions into a single sentence. On the other hand, the main text only contains one table, while all other 27 tables are placed in the appendix and mentioned only by forward reference. This requires the readers to switch back and forth between the main text and appendices to understand and verify the paper, which makes the reading experience very terrible.

Overall, these presentation issues are not just only a writing issue for me. This unclear presentation makes it very difficult to evaluate the actual contribution and soundness of the paper. As I am required to wire explanations from different sections together and try to guess what the author means or why the conclusions are drawn from those experiments and explanations. As a result, I could only understand the paper at a high level, and found it very difficult to verify its technical claims or to understand what new insights the paper actually provides.

2. Unclear Practical Importance of the Directional Bias.

Using MSE or martingale nulls for controlled experiments is not a problem by itself. However, since the paper focuses on financial forecasting, it is important to show how the identified directional bias actually affects a meaningful downstream task.

In financial markets, returns are not necessarily symmetric, and a model predicting upward movements more often is not necessarily a bad thing. For example, if a market tends to rise over time, it may actually be reasonable for a model to learn this trend or pattern from its pretraining data. Moreover, the paper itself also shows that removing the bias often makes the forecasting performance worse on real financial data. So the main question will be: why should one consider this bias as a problem in the first place in financial data, and what is the actual benefit of removing it?

Although the paper does include some directional accuracy results in appendix (and the bias correction actually hurt the directional accuracy), most of the evaluation still focuses on MSE-based skill and synthetic null settings. A more direct connection to the real world finance forecasting task will actually be using realized returns, cross-sectional IC/RankIC, or a simple portfolio backtest under the real-world financial data. Without such evidence, it is difficult to understand whether the identified bias is actually harmful or useful to the real world financial data, and why finance practitioners should care about it and why it needs to be removed.

3. Limited Evaluation and Evidence of Finance-Specific Foundation Models.

Another concern is the choice of the evaluated models. The paper mainly focuses on evaluating general-purpose time-series foundation models, with FinCast being the only financial-specific model included. However, there are already foundation models specifically designed for financial time series, such as Delphyne[1] and Kronos[2], and they both show that general-purpose time-series foundation models do not necessarily perform well on financial forecasting tasks.

On the other hand, an interesting difference between general-purpose or financial-specific foundation models can already be observed from the experiments in the paper: Even without any correction, FinCast naturally does not show the same upward bias as several general-purpose models does. This raises an interesting question about whether financial pretraining will let the model have different directional biases. However, this observation is only based on a single model, and the different behavior of FinCast can also come from its model architecture or training objective design, so more evaluation will be needed to confirm this assumption.

Overall, if these general-purpose models are not particularly suitable for financial forecasting in the first place, it is unclear why studying their directional biases would be useful for practical financial forecasting.


[1] Ding, Xueying, Aakriti Mittal, and Achintya Gopal. "Delphyne: A pre-trained model for general and financial time series." arXiv preprint arXiv:2506.06288 (2025).

[2] Shi, Yu, et al. "Kronos: A foundation model for the language of financial markets." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 30. 2026.

Questions for authors

Q1: I would like to see a more direct connection between the directional bias and real financial forecasting tasks. Can the authors compare the original and corrected predictions using:

  • (a) realized returns from trading based on the predicted direction,
  • (b) cross-sectional IC/RankIC, and
  • (c) a simple portfolio backtest?

As the paper already evaluates 49 industry portfolios, I think at least some of these additional evaluations should be relatively straightforward using the existing predictions. I think these three experiments are important because they actually reflect different aspects of real financial forecasting.

First, “evaluating the actual returns from trading based on the predicted direction” is important because lower directional accuracy does not necessarily indicate lower profit. A model may still gain more profit if it predicts correctly only during the periods with larger market movements. Therefore, I think it is important to evaluate whether removing the bias actually improves or hurts the realized returns, rather than solely looking at the direction accuracy.

Second, the cross-sectional experiment is also important because a model can still perform useful relative ranking between assets even if it contains directional bias. For example, if a model has the same upward bias across different assets, this bias may be canceled out when comparing their relative predicted returns, so it may not affect the final ranking at all. Therefore, having a directional bias does not necessarily mean the model is less performant in practice.

Third, a portfolio backtest will make the cross-sectional experiment more realistic. In practice, the investors usually trade only the top-ranked subset of assets rather than treating every prediction equally. This scenario will also allow the authors to evaluate more practical financial metrics, such as Annual Return (AR) and Sharpe Ratio (SR), and can further compare whether the original predictions or the corrected ones are actually more useful in practice.

Q2: I would like to see more evaluations on financial foundation models. Specifically, can the author evaluate Kronos[1]? as its code and pretrained checkpoints are publicly available on GitHub and HuggingFace.

I think it would be more interesting to compare whether finance-specific foundation models show similar behaviors to the general-purpose ones. For example, do finance-specific models show different directional biases? Does the upward bias mainly appear in some general-purpose models, or can it also be observed in finance-specific foundation models?

Overall, I think these experiments would make the findings much more meaningful for financial forecasting. Otherwise, the paper should better clarify how well its conclusions can be applied to finance-specific foundation models.

11 comments
AuthorsOpenReview ↗

Response to Reviewer 9Ts7 (1 of 7): W1, presentation

To W1, Thank you for the concrete examples. We agree that the submitted text was too dense and used terms before defining them. We rewrote the abstract, the introduction and Sections 2 to 6.

(a) The abstract now opens with the problem and the finding. The sentence you quoted is gone from the abstract. The idea behind it now reads "Forecasting each history and its mirror image, with every step negated, isolates the part of a forecast that ignores the direction of the history."

(b) "Skill" and "departure" are now defined in the first paragraph of the introduction. Our comments use the following terms.

  • History (context in the paper): the past values given to a model. A window is a history together with the future values to be forecast.
  • Last value: the forecast that a series stays at its last observed value. For a random walk no forecast is better.
  • Departure: the gap between a model's forecast and the last value.
  • Skill: the share of the last value's squared error that a model removes. It is zero for the last value and negative for a model that does worse.
  • Upward bias: a model forecasts a rise more often than a fall when nothing in the history calls for it. Five of the eleven models do, and we call them the upward-biased models.
  • Sign-randomised copy: a real window in which every step keeps its size and gets a random sign. It has no drift, so the last value is again the best forecast.
  • Mirror: the same history with every step negated, so that rises become falls.
  • Even part: forecast a history and its mirror, and average the two departures. A forecast that follows the direction of the history flips when the history is flipped, so that part cancels in the average. What remains does not depend on the direction of the history.
  • Mirror correction: subtract the even part from the forecast.
  • Memory: a model forecasts a period well because that period was in its training data.

(c) We split many of the sentences that combined a method, several numbers and a test. The main text has three tables instead of one. They hold what each model forecasts on random walks (Table 1), the financial tasks (Table 2) and the correction (Table 3). Table 1 brings together results that the submission tabulated only in the appendix. Where the main text points to the appendix, it now states the result first.

(d) You wrote that the presentation kept you from verifying the claims and from seeing what is new. The introduction now lists five findings, summarised below.

  • Five models forecast a rise when there is nothing to predict. On random walks every model loses to the last value, and these five forecast a rise on 70 to 85 percent of the walks (Table 1).
  • A corpus with positive trends is enough to produce the bias. Two small models trained from scratch on such a corpus acquire it (Figure 4).
  • On real prices, bias and memory are credited as skill. On rising equities, removing the even part lowers skill for ten of the eleven models, for four of them by more than two standard errors (Table 28). No model's directional trades beat holding the market by two standard errors (Table 2). Four models forecast the 1996-2014 market in a way that points to memory (Figure 5 and Table 21).
  • The bias is not specific to finance, and covariates can add their own. It appears on sign-randomised copies of electricity, traffic and weather windows, and an unrelated covariate that rises pulls the forecasts of two models upward (Appendices M and N).
  • The mirror correction is a diagnostic. It is guaranteed to lower error only when a series and its mirror image, past and future together, are equally likely. Mirror augmentation during fine-tuning removes most of Chronos-Bolt's bias without loss on ETTh1 (Tables 3 and 30).
AuthorsOpenReview ↗

Response to Reviewer 9Ts7 (2 of 7): W2, why the bias matters

To W2, Thank you for asking why this bias should count as a problem in financial data and what removing it gains. The bias is a problem for evaluation, and removing it brings no gain in returns. The revision now says both.

You are right that an upward bias is not harmful in itself. On a rising market a forecast that always rises earns money, and the revised introduction says so. A model may also learn the bias from rising pretraining data, as you suggest. Our corpus experiment shows that this can happen, since two small models trained on a corpus with positive trends acquire it (Figure 4).

But the bias does not depend on the series. The five upward-biased models forecast a rise on 70 to 85 percent of random walks, where nothing rises. Four of them keep the bias on sign-randomised copies of real equity windows (Table T1, column 3). Chronos-2 is the exception. Its departure on the copies is close to zero, so its upward forecasts on real windows cannot be counted as this bias. What the bias is worth then depends on the market.

(a) Where nothing rises, the bias is error. On random walks and on sign-randomised copies the last value is the best forecast, so a systematic upward departure is a systematic error (Tables 1 and 3).

(b) Where the market fell, skill is lower. On 1996-2014 windows the skill of all five models is lower where the market fell than where it rose. For Chronos-T5 it is +0.219 against −0.151, and for Moirai-1.1 +0.171 against −0.277 (Table 20). For these two models removing the even part raises skill where the market fell (+0.064 ± 0.050 and +0.208 ± 0.062). For the other three it does not (Section 5). Chronos-2's even part points down on those windows, and Chronos-Bolt's and TiRex's forecasts of those years point to memory.

(c) Where the market rose, a direction score rewards forecasting a rise. On 900 equity windows from 2017-2026 the five models call the direction of the next 128 days correctly 59 to 66 percent of the time. A forecast that always rises, which uses no model, is right 69 percent of the time (Table 23). Across the eleven models the score follows how often a model forecasts a rise, with a correlation of 0.96. You noted that the correction lowers directional accuracy. It does for all five, for the same reason. Section 5 now makes the comparison with always rising in the main text.

A good score on rising prices therefore does not show that a model has learned anything from the history it is given. Section 7 lists the checks to run before a score is read as skill. A model should be compared with the last value, a fitted drift and always rising, scored on sign-randomised copies, backtested on years after the training data, and tested with an unrelated covariate.

On the benefit of removing the bias, you observed that the correction often lowers performance on real financial data. The new backtests agree (see our replies to Q1). The revision therefore presents the mirror correction as a diagnostic that measures how much of a gain is the same whether the history rose or fell. Three models have point estimates above the last value on 2017-2026 equities, Chronos-T5, Chronos-Bolt and TiRex, with skill of 0.05, 0.13 and 0.12. The correction removes all or most of it and leaves −0.02, 0.05 and −0.07 (Table 35).

Table T1 (from Tables 1, 16, 23 and 35 of the revision, results already in the submission). The five upward-biased models at 128 steps. Column 2 is the share of forecasts that rise on random walks. Column 3 is the mean departure on sign-randomised copies of the 900 equity windows, in units of the daily return standard deviation of the history. Column 4 is the same on the raw windows. Column 5 is the share of forecasts that rise on the raw windows, and column 6 the percentage of the raw windows on which the forecast direction was right. Values after ± are standard errors.

ModelRises, random walksDeparture, copiesDeparture, rawRises, rawDirection right (%)
Chronos-T50.85+1.46 ± 0.08+2.65 ± 0.250.8461 ± 5
Chronos-Bolt0.79+1.37 ± 0.13+3.72 ± 0.420.8866 ± 5
Chronos-20.70+0.06 ± 0.08+1.93 ± 0.300.8564 ± 5
TiRex0.85+1.79 ± 0.11+2.19 ± 0.400.7866 ± 4
Moirai-1.10.73+2.32 ± 0.18+2.64 ± 0.320.7059 ± 3
Always rising----69 ± 6
AuthorsOpenReview ↗

Response to Reviewer 9Ts7 (3 of 7): Q1, the three financial evaluations

To Q1, Thank you for proposing the three evaluations and for explaining what each one measures. We ran all three with the raw and the mirror-corrected forecasts (Section 5, Table 2, Appendix K).

Twelve models, the eleven of the paper and Kronos-small, each forecast the market and the 49 industry portfolios on 114 dates over 2017-2026, 21 trading days apart. We trade on the 21-day forecast and hold for 21 days. Table T2 gives the Sharpe ratios and the RankIC.

(a) Trading on the predicted direction. Each asset is held long or short by the sign of its forecast. Holding the market earned 11.4 percent a year with a Sharpe ratio of 0.66. The best direction strategy, TiRex's, earned 7.5 percent a year with a Sharpe ratio of 0.70. No model beats holding the market by two standard errors. After seeing the results we also reran ten of the models on dates moved by 7 and by 14 trading days (a post hoc analysis). Averaged over the three schedules, none of the ten is above holding the market even in point estimates.

(b) IC and RankIC. On each date we correlate forecast and realised returns across the 49 industries. Only TiRex has a RankIC above zero after a Holm correction for testing eleven models (0.049 ± 0.018). The Pearson IC is above zero by two standard errors for TiRex (0.060 ± 0.017) and TimesFM-2.0 (0.050 ± 0.017).

(c) Portfolio backtest. The portfolio buys the ten industries with the highest forecasts and sells the ten with the lowest. The best baseline is 12-1 momentum, with a Sharpe ratio of 0.35. No model's portfolio beats it by two standard errors. In point estimates only TiRex's is above it (0.94). Annual returns are in Table T3 of our next comment. Volatility, drawdown, turnover, the Sharpe ratio net of costs and the three-factor alpha are in Table 32.

With 114 dates the standard error is 0.15 to 0.72 on a direction strategy's margin over holding the market and 0.25 to 0.32 on a long-short Sharpe ratio (Appendix K). The backtests detect no skill on these tasks apart from the Pearson IC of TiRex and TimesFM-2.0 and TiRex's ranking signal.

Table T2 (new runs, Table 2 of the revision without its three-schedule column). 2017-2026, raw → mirror-corrected. The direction strategy holds each asset long or short by the sign of its forecast. The long-short portfolio holds the top ten industries against the bottom ten. RankIC is the mean rank correlation of forecast and realised returns across industries, with its standard error. The last column is the raw direction Sharpe ratio on 1996-2014, years that public training corpora contain.

ModelDirection SharpeRankICLong-short SharpeDirection Sharpe, 1996-2014
Chronos-T5+0.18 → −0.18+0.013 ± 0.015 → +0.013+0.26 → +0.33+0.91
Chronos-Bolt+0.36 → +0.29+0.007 ± 0.021 → +0.009+0.02 → −0.12+1.59
Chronos-2+0.33 → +0.22+0.019 ± 0.016 → +0.004+0.27 → −0.07+0.77
TiRex+0.70 → +0.34+0.049 ± 0.018 → +0.034+0.94 → +0.56+1.78
Moirai-1.1+0.28 → −0.04+0.008 ± 0.014 → −0.000+0.29 → +0.02+0.27
Moirai-2.0+0.00 → +0.31+0.018 ± 0.020 → +0.019−0.05 → −0.04+1.46
TimesFM-2.0+0.34 → +0.09+0.039 ± 0.017 → +0.019+0.28 → −0.01+2.31
TimesFM-2.5+0.25 → +0.50+0.018 ± 0.017 → +0.021−0.04 → +0.01+0.41
Time-MoE+0.03 → −0.23−0.024 ± 0.020 → −0.032−0.28 → −0.50+0.77
Sundial−0.17 → −0.36−0.017 ± 0.018 → −0.058−0.19 → −0.98−0.02
FinCast−0.16 → −0.33−0.004 ± 0.014 → +0.004−0.12 → −0.06−0.07
Kronos-small+0.30 → −0.03−0.006 ± 0.017 → −0.017−0.13 → −0.10−0.22
12-1 momentum+0.21+0.026 ± 0.018+0.35+0.43
Holding the market+0.66--+0.51

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response

We thank the three reviewers. They find that the martingale setting with a known optimum gives a clean way to separate forecasting skill from directional bias (u2x9-S1), and that the phenomenon is interesting and relevant (EsHt-S1). They also credit the controls that separate confounding factors (EsHt-S2), the evidence across eleven models (u2x9-S2, 9Ts7-S1), and the mirror-based analysis (u2x9-S3).

The reviews raise three shared questions. The first concerns the practical importance of the bias and of its correction (9Ts7-W2, 9Ts7-Q1, u2x9-W1, u2x9-W3, u2x9-W4, u2x9-Q2, EsHt-W2, EsHt-Q1). The second concerns how far the findings reach, across models, domains and inputs, and how strong the evidence on memory is (9Ts7-W3, 9Ts7-Q2, u2x9-W2, u2x9-W5, u2x9-Q1, EsHt-W3, EsHt-Q2). The third concerns the presentation and the weight given to the theory (9Ts7-W1, EsHt-W1, EsHt-Q3).

The last value is the forecast that a series stays where it is, which is the best forecast for a random walk. The mirror correction forecasts a history and the same history with every step negated, and removes the part that the two forecasts share.

(1) Bias and memory can both pass for skill, and the correction does not raise returns. (a) Following 9Ts7-Q1, we traded on each model's forecasts of the market and 49 industry portfolios every 21 trading days over 2017-2026 (Table 2, shown as Table T2 in our reply to Reviewer 9Ts7). Holding the market had a Sharpe ratio of 0.66.

  • No strategy that goes long or short by the sign of a model's forecast beats it by two standard errors.
  • No portfolio of top against bottom industries beats 12-1 momentum by two standard errors.
  • The one exception is TiRex, which ranks the industries better than chance after a correction for testing eleven models. A post hoc check finds this signal schedule-dependent by a narrow margin.

(b) Of 96 paired comparisons between raw and mirror-corrected forecasts, 13 differ by more than two standard errors before any correction for multiple comparisons, and 12 of these favour the raw forecasts. The revision therefore presents the mirror correction as a diagnostic. It measures how much of a gain is the same whether the history rose or fell.

(c) Four models forecast the market's moves over 1996-2014 far better than rules read from the same histories, and public training corpora contain those years. Their direction strategies have Sharpe ratios of 1.46 to 2.31 on those years and 0.00 to 0.70 on 2017-2026 (0.35 to 0.62 averaged over three rebalancing schedules, a post hoc analysis). The revision asks that an upward bias and remembered history both be ruled out before a score is read as skill (abstract and Section 7).

(2) How far the findings reach. (a) Kronos shows no upward bias on random walks at the 128-step horizon. With 512-step histories it has the largest mean upward departure of any model on sign-randomised equity copies (post hoc). It loses to the last value and beats neither holding the market nor momentum (Tables T4 and T5 in our replies to Reviewer 9Ts7).

(b) For Chronos-Bolt, TiRex and Moirai-2.0 the mirror and copy checks point to memory. For TimesFM-2.0 the calendar-shift check gives a clear signal (Table U2 in our reply to Reviewer u2x9).

(c) We made copies of electricity, traffic, weather and ETTh1 windows in which every step keeps its size and gets a random sign, so that no direction can be predicted. The five models with an upward bias forecast a rise on 55 to 73 percent of these copies (Appendix N).

(d) An unrelated covariate that rises instead of falling raises the share of rising forecasts by 0.40 for Chronos-2 and 0.54 for Moirai-2.0 (Table E1 in our reply to Reviewer EsHt).

(3) The text is rewritten, and the theory is presented as supporting analysis. (a) We rewrote the abstract, the introduction and Sections 2 to 6, defined the main terms in the introduction, and raised the number of tables in the main text from one to three. The title, the section structure and the original experiments are unchanged.

(b) Sections 1 and 2 describe the two propositions as standard results used as measuring tools, and Section 7 ends with checks that need only forecasts.

The revised PDF has been uploaded. New or changed text is in the reviewer's colour (dark red for 9Ts7, blue for u2x9, green for EsHt), reworded text is black with a bar, and margin labels name the comment each change answers. Section and table numbers refer to the revision.

AuthorsOpenReview ↗

Summary of changes in the rebuttal revision PDF

We thank all three reviewers again for their time and comments. We have uploaded the revised PDF and replied to each review with the new experiments and analyses. New or changed content is in the reviewer's colour, and a label in the right margin names the review comment it answers. The colours are dark red for 9Ts7, blue for u2x9 and green for EsHt. Text that is only reworded stays black, with a bar or a label in the margin. The main changes are listed below.

  • Abstract and Section 1, rewritten, with the main terms defined where they first appear, a new paragraph on why the bias matters for evaluation and a new fourth finding on series outside finance, covariates and Kronos (9Ts7-W1, 9Ts7-W2, 9Ts7-W3, 9Ts7-Q1, 9Ts7-Q2, u2x9-W2, u2x9-W3, u2x9-W4, u2x9-Q1, u2x9-Q2, EsHt-Q1, EsHt-Q2, EsHt-Q3)
  • Sections 2 to 6, rewritten, with the two propositions presented as standard results that serve as measuring tools (9Ts7-W1, EsHt-W1, EsHt-Q3)
  • Table 1, new in the main text, with the null results that the submission tabulated only in the appendix (9Ts7-W1)
  • Figures 6 and 10, moved from Sections 3 and 5 to Appendices D.1 and F.1 (9Ts7-W1)
  • Section 3, Kronos on the null processes (9Ts7-W3, 9Ts7-Q2)
  • Section 4, a new paragraph on covariates (EsHt-Q2, EsHt-W3)
  • Section 5 and Table 2, new backtests over 2017-2026 that trade on the forecasts of the market and 49 industry portfolios, with RankIC and with holding the market and 12-1 momentum as benchmarks (9Ts7-Q1, 9Ts7-W2)
  • Section 5, the evidence on remembered years, from the calendar-shift check and from direction Sharpe ratios on 1996-2014 against 2017-2026 (9Ts7-W2, u2x9-W2)
  • Section 5, sign-randomised copies of electricity, traffic, weather and ETTh1 series, and how the copies separate bias from a legitimate drift (u2x9-Q1, u2x9-Q2, u2x9-W5, u2x9-W2)
  • Section 6, the mirror correction presented as a diagnostic that does not raise returns in the backtests (u2x9-W1, u2x9-W4)
  • Section 7, rewritten with two new checks, a backtest on years after the training data and a test with an unrelated covariate (9Ts7-W2, EsHt-Q1, EsHt-W2)
  • New Appendix K, the financial evaluation of raw and corrected forecasts, with IC and RankIC, three rebalancing schedules, paired comparisons and post hoc robustness checks (9Ts7-Q1, 9Ts7-W2)
  • New Appendix L, Kronos on the null processes, on equity windows and in the backtest, with three history lengths (9Ts7-Q2, 9Ts7-W3)
  • New Appendix M, covariates and multivariate inputs on the null, on the financial task and with real covariates (EsHt-Q2, EsHt-W3)
  • New Appendix N, non-financial series and post hoc checks on remembered weather data (u2x9-Q1, u2x9-W5, u2x9-W2)
  • Appendices I and J, implications and limitations updated for the new experiments, and related work on Kronos and Delphyne (9Ts7-W1, 9Ts7-Q1, 9Ts7-Q2, 9Ts7-W3, u2x9-Q1, u2x9-W2, EsHt-Q1, EsHt-Q2, EsHt-W2)

We hope the new experiments and analyses address your concerns, and we would be glad to answer any further questions.

AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their careful reviews and specific suggestions, and for reading our responses. Your comments showed us where the paper needed more evidence, mainly on financial tasks and outside finance, and where the checks for the bias needed a clearer explanation. We revised the paper accordingly. Having now gone through every reviewer's comments, we summarize once more what the revision adds.

  1. Financial evaluations. We added backtests over 2017-2026 that trade on each model's forecasts of the market and 49 industry portfolios. They compare raw and mirror-corrected forecasts in three ways, trading on the predicted direction, IC and RankIC, and a portfolio of the top against the bottom industries. Post hoc checks repeat them on two further rebalancing schedules and vary the holding period, portfolio size and cost. (Section 5, Table 2 and Appendix K)

  2. Practical guidance and the role of the correction. We now present the mirror correction as a diagnostic and state when it is guaranteed to lower risk. Section 7 ends with four checks that need only forecasts, two of them new, a backtest on years after the training data and a test with an unrelated covariate. We also compared the corpus of our controlled experiment with public pretraining corpora. It rises no more often than the Chronos datasets and LOTSA, and it was enough to give our small models an upward bias. (Sections 1, 6 and 7, Appendices E.4 and I)

  3. Finance-specific models. We evaluated Kronos-small and Kronos-base on random walks, on equity windows and their sign-randomised copies, and in the backtest. Kronos-small was also run with three history lengths. (Sections 1 and 3, Appendix L)

  4. Non-financial domains. We ran the eleven models on electricity, traffic, weather, ETTh1 and influenza-like-illness series and on their sign-randomised copies, with post hoc checks on remembered weather data. (Section 5, Appendix N)

  5. Memory and drift. We compared the direction strategies on 1996-2014, years that public training corpora contain, with those on 2017-2026. Section 5 also states how sign-randomised copies separate bias from a legitimate drift. (Section 5, Appendices F.5 and K)

  6. Covariates and multivariate forecasting. For the four models that accept further inputs, we tested covariates and joint inputs on the null processes and on the financial task, and with the real covariates of ETTh1 and Weather. (Section 4, Appendix M)

  7. Presentation and theory. We rewrote the abstract, the introduction and Sections 2 to 6 and defined the main terms where they first appear. The main text now has three tables instead of one, and it presents the two propositions as standard results used as measuring tools. (Sections 1 to 6)

The main text gives the most important new results, and the sections and appendices above give the full settings and results. New or changed content is in each reviewer's colour, with a margin label naming the comment it answers. Thank you again for your time and comments.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.