acceptodds
Desk rejected at ICLR 2027

DU-TTT: Decision-Utility Test-Time Training for Portfolio Construction with Delayed Feedback

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

A deployed return forecaster must adapt online to non-stationary environments. Standard test-time training (TTT) strategies fail here for two reasons: misalignment, as mean-squared error ignores decision boundaries and portfolio mechanics, and delayed, constrained feedback, as labels arrive days later under strict leverage and fee constraints. To address these challenges, we propose Decision-Utility Test-Time Training (), which adapts a lightweight affine head on a frozen feature extractor directly using downstream portfolio utility. buffers historical decision contexts, evaluates matured -day returns, and updates the adaptation head via a covariance-weighted utility gradient. To focus feedback where it matters most, incorporates a soft top- selection mechanism, concentrating adaptation on the active trading margin while dynamically adjusting step size based on forecast dispersion. Evaluated across 109 equity universes over two distinct market regimes, consistently outperforms standard MSE-based adaptation in both portfolio growth and Sharpe ratio. Our analysis reveals that achieves its primary advantage by suppressing unrewarded turnover and trading friction. Furthermore, ablation studies confirm the necessity of both the margin-focused selection and dispersion-scaled step sizes. By directly aligning test-time adaptation with financial utility, offers a robust framework for real-world algorithmic trading under distribution shift.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Quant Investment, AI, Agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6hd1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an online method for adapting return forecasts when feedback arrives with a delay. It updates a small prediction head using portfolio utility and focuses learning on assets near the selection boundary. Experiments with several forecasting models show better trading performance and lower turnover than the reported baselines.

Critical strengths
  1. The paper addresses a relevant practical problem. Adapting only the prediction head makes the method simple to use with different forecasting models.

  2. The evaluation considers trading costs and turnover. This makes the results more useful than a comparison based only on forecast accuracy.

Critical weaknesses

My two main concerns are:

  1. I am not convinced that the utility objective is the main reason for the improvement. In the ablation, a squared-error update already performs well when given the proposed step rule. The paper should explain more carefully what the utility direction adds.
  2. The step-size derivation needs clarification. Forecast dispersion depends on the predictions, but the stated derivative treats it as fixed. The authors should explain whether this is an intentional approximation and whether it affects the results.
Questions for authors
  1. Does the code treat forecast dispersion as fixed? Please clarify how this agrees with Equation (9).
  2. Can you compare all baselines on the held-out period without retuning, including the squared-error update with DU-TTT’s step rule? This would help clarify whether the reported gains hold on new data.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 6hd1

We thank the reviewer for the questions on what the utility objective contributes (W1), on the dispersion term in the step size (W2/Q1), and on the held-out comparison (Q2).

Summary.

  • W1. The step rule produces most of the gain over affine MSE in folds 2–3 and about half in fold 1. The utility direction adds lower turnover (26–43% less at the same step), which turns into a net gain as the fee rises. Lines 407–408 will be rewritten to say this.
  • W2/Q1. The code holds the dispersion fixed, as an intentional approximation, and only the step length depends on it. Across 0–50 bps, the exact derivative changes cell SR by at most 0.021, and every interval includes zero.
  • Q2. On the held-out fold 3, DU-TTT leads the squared-error update with its step rule at 50 bps and is ahead of every decision-aligned and ranking objective (intervals exclude zero except ListMLE). RLS and OGD are ahead within noise, and the frozen forecast is ahead.

Notation. Cell SR: mean net Sharpe over the 327 cells at 25 bps. Brackets: 95% intervals of DU-TTT minus the named rule. [new]: run for this response, no re-tuning.

W1. What the utility direction adds.

cell SR, 25 bpsFold 1Fold 2Fold 3
Affine MSE (own step)0.2810.0300.072
Squared-error direction, DU-TTT step0.5150.6370.461 [new]
DU-TTT0.7120.7120.562

The step adds +0.234 / +0.607 / +0.389, and the direction adds +0.197 / +0.075 / +0.101. The direction's contribution depends on the fee [new]:

DU-TTT − sq. error, same step0 bps25 bps50 bpsturnover per rebalance
Fold 1+0.050 [−0.053, +0.150]+0.197 [+0.086, +0.305]+0.341 [+0.226, +0.458]0.305 vs 0.529
Fold 2−0.016 [−0.133, +0.098]+0.075 [−0.045, +0.192]+0.166 [+0.045, +0.285]0.233 vs 0.411
Fold 3+0.041 [−0.086, +0.165]+0.101 [−0.031, +0.231]+0.161 [+0.023, +0.291]0.414 vs 0.563

At 0 bps the two directions are tied. At 50 bps the utility direction wins in all three folds. The lower turnover follows from the construction. Qs1=0Q_s\mathbf 1=0 removes the common-level component, which squared error still moves along (Appendix A.1, A.4). The covariance weighting penalizes only errors that move weight between assets (Eq. 6). [new] The same ordering holds with a 130/30 long–short decision layer: +0.255 [+0.143, +0.369] / +0.146 [+0.028, +0.259] in folds 1 / 2.

W2/Q1. Dispersion in Eq. (9). Yes. The code holds μ^ˉs\bar{\hat\mu}_s and ςs\varsigma_s fixed, and Eq. (9) is that derivative; we will say so. Since gsK⊥1g^K_s\perp\mathbf 1, the exact derivative is gsK,exact=(I−1nzszs⊤)gsK,g^{K,\mathrm{exact}}_s=\Bigl(I-\tfrac1n z_s z_s^\top\Bigr)g^K_s , which removes the component along zsz_s. That component only rescales μ^s\hat\mu_s and leaves the soft top-KK membership unchanged. Finite differences confirm the formula.

DU-TTT uses gsKg^K_s only for its length (the direction is gsbfg^{\mathrm{bf}}_s), so the approximation can change only the step size. [new] The exact length is a median of 1.064 / 1.071 times the fixed one (folds 1 / 2). Re-running with the exact length gives fixed − exact = +0.010 [−0.019, +0.040], +0.018 [−0.005, +0.042] and −0.001 [−0.028, +0.024] in folds 1 / 2 / 3 at 25 bps. A step about 7% longer is well inside Table 5, where multipliers of 0.5–2× change cell SR by at most 0.097.

Q2. Held-out period, no re-tuning. Fold 3 (88 days) was not used in design and was evaluated once. Every rule runs at its fold-2 selection.

Fold 3cell SRDU-TTT − row
Frozen / static rescaling / Hedge0.926 / 0.931 / 0.923−0.363 / −0.368 / −0.361
Affine RLS0.702−0.140 [−0.297, +0.018]
OGD (absolute error)0.597−0.035 [−0.173, +0.103]
Sq.-error dir., DU-TTT step [new]0.461+0.101 [−0.031, +0.231]
ListMLE0.431+0.132 [−0.008, +0.270]
Online Newton step0.397+0.165 [+0.032, +0.298]
LambdaRank-IC / Soft Spearman0.270 / 0.192+0.292 / +0.370
Quantile tracking / SPO+ (unit)0.142 / 0.167+0.420 / +0.396
SPO+ / Fenchel–Young0.261 / 0.160+0.302 / +0.403
Affine MSE0.072+0.490 [+0.349, +0.634]
DU-TTT0.562reference
  • Against the squared-error update with DU-TTT's step, DU-TTT leads by +0.161 [+0.023, +0.291] at 50 bps. On the aggregate strategy, its Sharpe is 1.574 against 1.440, at turnover 16.6 against 22.1 per year.
  • DU-TTT also leads its own ablation without the dispersion-scaled step by +0.158 [+0.044, +0.272].
  • In fold 3 the frozen forecast is already strong (0.926, against 0.157 / 0.237 in folds 1 / 2), and no online rule beats it.
  • Weighting the three periods equally, DU-TTT leads RLS by +0.179 [+0.097, +0.255] and the frozen forecast by +0.222 [+0.132, +0.315]. Folds 1–2 informed selection, so this is not a held-out figure.

We will add this table to Section 5 and state in the limitations when adapting does not help.

AuthorsOpenReview ↗

Summary of rebuttal revisions

ConcernRevisionWhere
W1: is the utility objective the reason?The overstating sentence is replaced: the step is a generic lever, and the direction adds 0.197/0.075 at equal step. Fee-dependence table: the direction ties at 0 bps and wins at 50 bps in all three periods, with 26–43% lower turnover.Sec. 5.2; App. H.3, Table 13
W2/Q1: dispersion treated as fixedBefore Eq. (9): "with μ^ˉs,ςs\bar{\hat\mu}_s,\varsigma_s fixed". The exact derivative (I−1nzszs⊤)gsK(I-\tfrac1n z_sz_s^\top)g^K_s is given; using it changes cell Sharpe by at most 0.021.Sec. 4.3; App. H.4
Q2: all baselines on the held-out period, including squared error with the DU-TTT stepTable 11 includes the new squared-error-at-DU-TTT-step row (0.461; aggregate Sharpe 1.440 vs 1.574 for DU-TTT).App. G, Table 11

Hope the revision and clarify in rebuttal can solve your concerns and consider increase the rating of the paper. If you have any continue concerns please let us know.

AuthorsOpenReview ↗

Additional results about what the utility update adds beyond the step rule

Dear Reviewer 6hd1,

Your observation holds on the 88-day test. Giving squared error the dispersion-scaled step rule recovers much of the improvement. Each method kept its own fold-2 settings (Appendix G, Table 11).

Held-out policy, at 25 basis pointsMean cell SharpeAnnualized portfolio turnover
Ordinary affine squared error0.07239.0×
Squared error with DU-TTT’s step rule0.46122.1×
DU-TTT0.56216.6×

Against squared error using the same step rule, DU-TTT’s cell Sharpe difference was +0.101 (95% interval −0.031 to +0.231). Rescoring the same trades at 50 basis points widened that gap to +0.161 (95% interval +0.023 to +0.291, Table 13). On the combined portfolio, annualized fee savings of 1.37 percentage points offset a 1.28-point gross-return shortfall, leaving +0.09 points in net return.

Eq. (9) holds forecast dispersion fixed to compute the step length. Differentiating through dispersion changed held-out mean cell Sharpe by 0.001 at 25 basis points (Appendix H.4).

If this clarifies the roles of the step rule and utility update,we would be grateful for any further comments, and we would ask you to consider whether the revised evidence solved your concerns also increase the overall score.

Best

The authors

Reviewer_Q275OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The theme of the work is important with growing interest. The paper studies how stock return prediction models can adjust to changing markets when actual returns become available after a delay. More accurate forecasts do not always lead to better investment decisions, and frequent trading adds costs. DU-TTT keeps the base model fixed and updates only a small calibration layer using actual returns and portfolio risk. It focuses on updates on stocks near the cutoff for portfolio selection and adjusts the size of each update. Its main contributions are linking online calibration under delayed feedback to portfolio decisions and introducing methods to control where and how much to update. The authors evaluate DU-TTT through historical backtests on 109 groups of U.S. stocks, three forecast horizons, and four prediction models. They also conduct ablation studies and test the method on a previously unused period. Compared with calibration based on prediction errors, DU-TTT generally reduces portfolio turnover and improves returns after trading costs and risk-adjusted performance.

Critical strengths

The paper presents a lightweight online calibration method that accounts for portfolio decisions and delayed feedback. It keeps the existing prediction model and portfolio optimizer unchanged and updates only two parameters per stock. Returns observed later are linked to the original decision context. This allows existing models to adapt after deployment while considering how forecast changes affect portfolio holdings and risk. Backtest results showing lower turnover and better performance after trading costs support this design.

Critical weaknesses

I have a couple of concerns:

  1. The method’s ability to adjust holdings effectively when position limits are reached remains unclear. Once a stock reaches its limit, increasing its predicted return cannot increase the holding further. The method then needs to adjust the allocation among other stocks. Its updates may keep favoring the stock already at its limit, but the paper does not sufficiently examine how this affects the allocation of the remaining capital.

  2. The main test data were also used to choose the final method. This creates a risk of overfitting those data, so the reported gains may be too optimistic. The additional independent test covers only 88 trading days and does not provide clear statistical evidence that the combined portfolio earns higher returns after trading costs.

Questions for authors

I have a couple of questions/clarifications that I need the authors to elaborate on:

  1. When a stock reaches its position limit, could you use the existing backtests to show how later DU-TTT updates change the allocations to other stocks? Do these updates improve portfolio performance in terms of returns and risk?

  2. The 88-day independent test mainly compares DU-TTT with MSE calibration and its own ablation variants. Could you use the same test data to add comparisons with other main baselines, such as RLS and calibration based on absolute error?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Q275

We thank the reviewer for the questions on updates at the position limit (W1/Q1) and on selection and the held-out test (W2/Q2).

Summary.

  • Position limit. An update to a stock at its limit cannot change how the remaining capital is split among the other stocks. DU-TTT's updates to the other stocks do reallocate capital. That reallocation earns positive return in fold 1, is neutral in fold 2, and loses in fold 3, the period in which the frozen forecast is strongest. It lowers portfolio volatility in all three folds.
  • Held-out test. Folds 1–2 informed the final choice, so fold 3 is the held-out test. On fold 3, DU-TTT is ahead of affine MSE and of the decision-aligned and ranking objectives. RLS and absolute-error calibration are ahead within noise, and the frozen forecast is ahead. Over all three periods, DU-TTT leads both.

How to read the numbers. Cell SR is the mean net Sharpe ratio over the 327 universe–horizon portfolios at 25 bps. Brackets are 95% intervals; if an interval contains zero, the two methods are not separated. [new] marks a run made for this response, with no re-tuning.

W1/Q1. Updates at the position limit.

(a) Mechanism. The head is per stock (μ^i=aiμi0+bi\hat\mu_i=a_i\mu^0_i+b_i), so an update to a capped stock changes no other forecast. While stock jj is at the limit, wj=wmax⁡w_j=w_{\max}. The remaining 1−wmax⁡1-w_{\max} is allocated from the other stocks' forecasts and Σ\Sigma, independently of μ^j\hat\mu_j (Appendix A.3).

(b) How often [new]. The limit binds in 99.9% / 99.8% of decisions (folds 1 / 2). On average, 1.96 of the 3.15 held stocks are at the limit. Only 6.6% / 7.3% of the update mass pushes a capped stock higher; most of the update goes to the stocks whose selection is open.

(c) Effect on the other stocks [new]. At every decision with a stock at the limit, we compared DU-TTT's target with the frozen forecast's target. We measured the return to the next rebalance, before fees, of the weight moved among the non-capped stocks, and the ex-ante volatility of each target.

Fold 1Fold 2Fold 3
Reallocation among non-capped stocks, pp/yr+4.68 [+2.58, +6.66]+0.21 [−2.34, +2.74]−8.47 [−12.72, −3.99]
Volatility %, DU-TTT / frozen21.3 / 24.423.6 / 25.325.4 / 29.3

So later updates do move capital among the other stocks, through those stocks' own forecasts. This adds return in fold 1, is neutral in fold 2, and loses in fold 3, where the frozen forecast is strong. It lowers risk in every fold. Counting all weight moved relative to the frozen target, DU-TTT's point estimate is above affine MSE's in each fold.

(d) Removing pushes on capped stocks [new]. We dropped every update component that raises a capped stock. This raises turnover by 12–24% and does not improve cell Sharpe. At 50 bps it lowers fold-1 cell SR (+0.104 [+0.041, +0.165] in favour of DU-TTT). These pushes keep capped positions held instead of rotating them out and back in.

W2/Q2. Selection and the held-out test. As stated in Appendix E (lines 1311–1318), the final configuration was chosen among twelve variants on folds 1–2. We will state this in Section 5. Fold 3 was not used in design and was evaluated once. Every rule runs at its fold-2 selection, with no re-tuning.

Fold 3 (88 days)cell SRDU-TTT − row
Frozen forecast0.926−0.363 [−0.534, −0.186]
Affine RLS0.702−0.140 [−0.297, +0.018]
OGD (absolute error)0.597−0.035 [−0.173, +0.103]
Online Newton step0.397+0.165 [+0.032, +0.298]
SPO+0.261+0.302 [+0.161, +0.438]
Affine MSE0.072+0.490 [+0.349, +0.634]
DU-TTT0.562reference

The other decision-aligned and ranking rules also score below DU-TTT (Fenchel–Young 0.160, ListMLE 0.431, LambdaRank-IC 0.270, Soft Spearman 0.192). RLS and OGD are ahead of DU-TTT within noise. The frozen forecast is ahead: fold 3 is a period in which it is already strong, with cell SR 0.926 against 0.157 / 0.237 in folds 1 / 2, and every online rule scores below it.

Weighting the three periods equally, DU-TTT leads RLS by +0.179 [+0.097, +0.255], OGD by +0.169 [+0.099, +0.239] and the frozen forecast by +0.222 [+0.132, +0.315]. Folds 1–2 informed selection, so this is not a held-out figure.

Statistical evidence. The reviewer is right that 88 days are too short to separate paired returns on a single portfolio but the topic and phenomenon is very valuable we will extend later to prove efficiency of the task. The excess return over affine MSE is +6.51 [−15.37, +28.39] pp/yr on fold 3. Because the window is short, the pre-registered fold-3 endpoint was the cell SR. It uses all 327 portfolios, and on that endpoint DU-TTT leads affine MSE by +0.490 [+0.349, +0.634]. It also leads DU-TTT without the dispersion-scaled step by +0.158 [+0.044, +0.272].

AuthorsOpenReview ↗

Summary of rebuttal revisions

ConcernRevisionWhere
W1/Q1: allocation to other stocks once a stock is cappedMechanism argument; frequency of the cap. New measurement of the reallocation among non-capped stocks: +4.68 / +0.21 / −8.47 pp/yr, with lower volatility in all folds. Cap-mask control.App. H.6, Table 15
W2: selection on the main test dataAlready stated in Appendix E (lines 1311–1318). Appendix G adds the all-baseline held-out table. The Section 5 sentence is deferred to the final version.App. E, G
Q2: other baselines (RLS, absolute error) on the held-out periodFull fold-3 table with every baseline, plus the three-period mean.App. G, Table 11

Hope the revision and clarify in rebuttal can solve your concerns and consider increase the rating of the paper. If you have any continue concerns please let us know.

AuthorsOpenReview ↗

Additional results about position limits and the held-out baselines

Dear Reviewer Q275,

When a stock stays at its 35% cap, updates to the other stocks can still change how the remaining capital is invested. On held-out decisions with a binding cap, 269 of 327 universe–horizon portfolios had lower average estimated volatility under DU-TTT than under frozen forecasts. In 174 of those 269, shifting capital among uncapped stocks made a negative gross return contribution relative to those stocks’ mean return (Appendix H.6).

Folds 1 and 2 informed selection. Fold 3 kept fold-2 settings fixed. The table gives combined-portfolio net return differences over those 88 days in percentage points (pp).

DU-TTT minus baseline, at 25 basis pointsAnnualized difference95% simultaneous interval
Recursive least squares−6.88 pp−27.47 to +13.72 pp
Absolute-error calibration−3.60 pp−21.00 to +13.81 pp

Both intervals span zero. Against ordinary affine squared error, DU-TTT improved mean cell Sharpe by 0.490 (95% interval 0.349 to 0.634) and cut annualized portfolio turnover from 39.0 to 16.6 times portfolio value (Appendix G).

If these comparisons answer your questions about the cap and the independent test, we would be grateful for any further comments, and we would ask you to consider whether the revised evidence solved your concerns also increase the overall score.

Best

The authors

Reviewer_6FszOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Decision-Utility Test-Time Training (DU-TTT), a method to adapt return forecasts from a frozen prediction model to changing market conditions using the latest available realized returns as feedback signals. Instead of updating the base model directly, the adaptation is through a linear transformation of the frozen model forecast.

Unlike existing online methods that minimize prediction error (e.g. MSE), the proposed method drives adaptation using portfolio utility and focuses updates on marginal assets near the portfolio selection boundary. The paper also provides mathematical justification.

Experiments are carried out across various. trading universe. Proposed method is compared to various common baselines.

Critical strengths
  • The paper addresses a critical real-world challenge in quantitative finance: adapting return forecasts to shifting market conditions under delayed feedback and trading fees without introducing future data leakage.
  • By tuning only two linear parameters per asset (multiplier and offset), the update is fast, adds minimal computation, and avoids overfitting to noisy financial returns.
  • Because it operates strictly on model outputs rather than internal weights, the method works as a plug-and-play layer across any frozen model.
  • Mathematical justifcaiton is provided.
Critical weaknesses
  • The paper is hard to follow because many symbols and terms are not defined or explained clearly. For example, w~t\tilde{w}_t and "cell" are used in Section 5 without proper definition, and Equation (7) appears out of context without clear explanation.
  • The framework only covers long-only portfolios (0≤w≤wmax⁡10 \le w \le w_{\max}\mathbf{1}) and does not evaluate long-short strategies.
  • The portfolio optimization problem in Equation (1) is too simple. Because it does not include transaction costs directly in the optimizer, it is unclear if the proposed method is still needed when using a realistic cost-aware portfolio optimizer.
Questions for authors

why are ranking objectives used as a baseline?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6Fsz

We thank the reviewer for the comments on notation (W1), long-only portfolios (W2), the cost term in Eq. (1) (W3) and the ranking baselines (Q).

Summary. (1) The flagged symbols are defined below and will be defined in the main text. (2) DU-TTT is not limited to long-only books: the update needs only an equality constraint on the decision. On a new 130/30 long–short layer, it keeps its lead in folds 1–2. (3) With a cost-aware optimizer in place of Eq. (1), every rule improves, and DU-TTT still has the highest cell Sharpe in folds 1–2. (4) Ranking objectives test whether forecast order alone is enough; for the three ranking losses it is not.

Notation. Cell SR: mean net Sharpe over the 327 cells at 25 bps. Brackets: 95% universe-cluster intervals of DU-TTT minus the named rule. [new]: run for this response at the validation-selected configuration, no re-tuning.

W1. Notation.

  • w~t\widetilde w_t (line 311) is the pre-trade holding at close tt: yesterday's post-trade weights after drifting with today's returns, w~t,i=wt−1,i(1+rt,i)/∑jwt−1,j(1+rt,j)\widetilde w_{t,i}=w_{t-1,i}(1+r_{t,i})/\sum_j w_{t-1,j}(1+r_{t,j}). The fee is charged on ∥wt−w~t∥1\|w_t-\widetilde w_t\|_1.
  • A cell is one universe–horizon pair (109 universes × h∈{1,5,20}h\in\{1,5,20\} = 327 cells), traded as an independent portfolio. The aggregate columns are one portfolio that holds all cells at equal capital.
  • Eq. (7). Figure 3 will move after the definitions, and Section 4.3 will open with the purpose. Under the cap, the portfolio holds about K=⌈1/wmax⁡⌉=3K=\lceil 1/w_{\max}\rceil=3 names, so the update must find the names near the selection boundary. The section then defines ysy_s as the soft top-KK membership of zs=(μ^s−μ^ˉs1)/ςsz_s=(\hat\mu_s-\bar{\hat\mu}_s\mathbf 1)/\varsigma_s, where νs\nu_s enforces ∑iys,i=K\sum_i y_{s,i}=K. Finally it defines δs,i=ys,i(1−ys,i)\delta_{s,i}=y_{s,i}(1-y_{s,i}), which is largest at the boundary. A notation table will be added.

W2. Long-only. The projection QsQ_s needs only an equality constraint. Appendix B (lines 1128–1134) gives Qs(B)Q^{(B)}_s for any set of linear equalities, including dollar-neutral books. [new] We replaced every rule's decision layer with a 130/30 book, {−0.10≤wi≤0.35, 1⊤w=1, ∥w∥1≤1.6}\{-0.10\le w_i\le 0.35,\ \mathbf 1^\top w=1,\ \|w\|_1\le 1.6\}. Results: first row of the table below.

W3. Cost term. The cost term reduces trading, but it does not replace the learning target; the two work together.

  • (a) Every calibrated family was also run with the same sixteen predicted-utility trade gates, a no-trade rule (Davis & Norman, 1990), with each gate selected on validation. With the gate, DU-TTT reaches aggregate Sharpe 1.281 / 0.952 at turnover 10.2 / 8.0 per year (folds 1 / 2). Affine MSE reaches 1.179 / 0.393 at 15.5 / 18.6. This table was left out of Section 5.3 and will be added.
  • (b) [new] We replaced Eq. (1) for every rule with max⁡w∈Wμ^⊤w−γ2w⊤Σw−κ∥w−w~∥1\max_{w\in\mathcal W}\hat\mu^\top w-\tfrac{\gamma}{2}w^\top\Sigma w-\kappa\|w-\widetilde w\|_1 (Boyd et al., 2017), where κ\kappa is the 25-bps fee amortized over the weekly holding period.
cell SR, folds 1 / 2 / 3FrozenAffine MSEAffine RLSSq.-error dir., DU-TTT stepDU-TTT
130/30 long–short−0.086 / −0.055 / 0.8090.053 / −0.165 / −0.2580.218 / 0.114 / 0.5770.257 / 0.468 / 0.2150.512 / 0.614 / 0.336
Cost-aware optimizer0.404 / 0.526 / 1.2220.584 / 0.186 / 0.1150.702 / 0.577 / 1.0580.659 / 0.716 / 0.5830.791 / 0.754 / 0.620
  • 130/30, folds 1–2. DU-TTT is ahead of every rule, and every interval excludes zero. The closest rule is the ablation that keeps DU-TTT's step but uses the squared-error direction: +0.255 [+0.143, +0.369] and +0.146 [+0.028, +0.259].
  • Cost-aware, folds 1–2. DU-TTT rises from 0.712 to 0.791 / 0.754. It leads RLS by +0.089 [−0.005, +0.184] / +0.178 [+0.063, +0.287] at 25 bps, and by +0.180 [+0.082, +0.281] / +0.251 [+0.132, +0.365] at 50 bps.
  • Fold 3. The frozen forecast and RLS score higher on both layers, as they do with Eq. (1). In this held-out period the frozen forecast is at its strongest (cell SR 0.926, against 0.157 / 0.237 in folds 1 / 2).

Q. Why ranking objectives? (1) Learning to rank is a standard approach to cross-sectional stock selection (Song et al., 2017; Poh et al., 2021). (2) Under the cap, the allocation is close to a top-KK selection, so a ranking loss tests whether getting the order right is enough. (3) DU-TTT's margin weight and step rule use the soft top-KK operator of the LML layer (Amos et al., 2019).

These three losses fall well short. ListMLE, LambdaRank-IC and Soft Spearman reach cell SR 0.052 / 0.329 / −0.030 (fold 1) and 0.038 / 0.027 / −0.077 (fold 2), at turnover 28–41 per year. DU-TTT reaches 0.712 / 0.712 at 10–14. They ignore covariance and reordering cost.

In Table 2, the rows labelled TAFAS, KKT-embedded DFL and IPMO are mislabelled: their numbers are those of ListMLE, LambdaRank-IC and Soft Spearman (Section 5, lines 306–307). The labels will be fixed.

AuthorsOpenReview ↗

Summary of rebuttal revisions

ConcernRevisionWhere
W1: undefined symbols (w~t\widetilde w_t, "cell"), Eq. (7)Notation table covering w~t\widetilde w_t, cell, aggregate and all symbols of Eqs. (3)–(9). The rebuttal points to the purpose sentence before Eq. (7) and to the definitions right after it.App. H.1
W2: long-only onlyNew 130/30 long–short decision layer for every rule. DU-TTT leads every rule in folds 1–2, with intervals excluding zero.App. H.5, Table 14
W3: no cost term in Eq. (1)(a) The matched trade-gate table is restored and cited in Sec. 5.3. (b) A new cost-aware optimizer is run for every rule. DU-TTT rises to 0.791/0.754 and stays first in folds 1–2.Sec. 5.3; App. H.2, H.5
Q: why ranking objectivesThe rebuttal states that the Table 2 rows labelled TAFAS / KKT-embedded DFL / IPMO are the ranking objectives. The relabelling is deferred to the final version.rebuttal only

Hope the revision and clarify in rebuttal can solve your concerns and consider increase the rating of the paper. If you have any continue concerns please let us know.

AuthorsOpenReview ↗

Additional results about does calibration still matter when the optimizer charges for trading?

Dear Reviewer 6Fsz,

With trading costs inside the optimizer, DU-TTT traded 0.318 times portfolio value per rebalance, versus 0.852 for affine squared-error calibration. Both used the same frozen forecasts and delayed labels. The table shows mean net Sharpe across 327 universe–horizon portfolios on the 88 held-out days, at a 25 basis point fee (Appendix H.5, Table 14).

OptimizerDU-TTTAffine squared error
Long-only with trading costs in the objective0.6200.115
130/30 long–short0.336−0.258

The advantage over squared error holds in both settings. Frozen forecasts and recursive least squares scored higher than DU-TTT in this period.

Eq. (7) focuses updates on stocks near the selection cutoff. Ranking losses test whether learning their order alone is enough. Appendix H.1 defines the symbols, and our earlier reply identifies the three ranking labels still to be corrected in Table 2.

If this addresses your concerns about costs, short positions and ranking, we would be grateful for any further comments, and we would ask you to consider whether the revised evidence solved your concerns also increase the overall score.

Best

The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Friendly Reminder to Reviewers

Dear Reviewers,

Thanks for your attention to this email. During the rebuttal phase, we extensively answered the questions raised by the reviewers and believe we have addressed the reviewers' concerns also make this interesting work better. This is a friendly reminder that the author-reviewer discussion period is coming to an end. We understand you are busy, but we would appreciate any additional feedback or suggestions you might have on our work. Your insights are valuable to us and DUTTT.

Sincerely,

The Authors

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

Related papers

Discussion (0)

Sign in to comment.