More historical data can improve estimation while introducing relationships that transfer poorly to current market conditions. We therefore examine how training quantity and historical selection interact with forecast horizon in intraday equity prediction. We deploy twelve configurations which combine four sampling and validation policies with 6, 12 and 24 month training quantities under a fixed cross-sectional Transformer and feature set. A monthly walk-forward experiment produces 264 refits and eight cumulative return forecasts from 5 to 120 minutes. Evaluation covers January 2025 through 20 October 2026 and uses cross-sectional Spearman information coefficients against future five-factor residual returns. Six-month recency sampling leads individual models at 5-15 minutes, while multi recency configurations lead at 20-30 minutes and uniform configurations at 60-120 minutes. Increasing training quantity does not uniformly improve accuracy while an equal-weight ensemble of all twelve models exceeds every individual mean at 15, 20 and 25 minutes, but not at the remaining horizons. These findings support evaluating historical selection jointly with forecast horizon without identifying a universal policy or a causal benefit from recency.
Figures & tables
Group
Count
Examples
Base inputs
72
Reversal, basket dislocations, risk exposures and context
Temporal transforms
160
Changes, moving means, history z-scores and acceleration
Nonlinear transforms
80
Squares, cubes, signed squares and tail excess
Context interactions
160
Products or damping by volatility, volume, liquidity and styles
Conditional signals
40
Reversal gated by volatility or liquidity regimes
Peer-relative signals
48
Deviations from liquidity, volatility, momentum or fitted-cluster peers
Table 1: Feature groups in the 572-input library. Extensions are deterministic transformations of shared anchor signals.
Family
Selected training months T
Pool months H
Final refit months
Short
6
12
8
Medium
12
24
14
Long
24
36
26
Table 2: Training-history families and eligible historical pools. Final refits add the latest two validation months to the selected training sample.
Figure 1: Monthly timeline for January 2025 forecasting. Above the calendar, C6 selects the newest six consecutive pool months (blue), whereas R6 selects six distinct months from the same pool (orange; an illustrative draw, not the recorded experimental draw). Both validate on November–December 2024 (grey). Below the calendar, each final-refit row identifies its six selected months plus the two validation months: these bars denote data membership, not when fitting occurs. A fresh model is refitted after validation and frozen before January test-month forecasts.
Horizon
Time-eligible rows
Observed total
Coverage (%)
5m
5,248,638
4,576,264
87.19
10m
5,248,638
4,573,370
87.13
15m
5,248,638
4,575,670
87.18
20m
5,248,638
4,577,204
87.21
25m
5,248,638
4,579,273
87.25
30m
5,248,638
4,600,076
87.64
Table 3: Endpoint-label coverage across all 22 test months; counted once, not twelve times across variants.
Figure 2: Monthly observed total-return label coverage for selected horizons. The denominator excludes horizons extending beyond the market close.
Setting
Value
Input features
572
Projected model width
32
Attention heads
4
Transformer layers
1
Feedforward activation
GELU
Normalization
Pre-norm
Table 4: Cross-sectional Transformer architecture and training settings shared by all twelve historical-selection configurations.
Model
Mean refit epochs
Median
Max
Selection epochs
C6
4.32
4.0
7
6.32
C12
5.23
5.5
11
7.23
C24
2.95
3.0
6
4.95
U6
4.14
4.0
9
6.14
U12
3.73
3.5
7
5.73
U24
3.36
3.0
7
5.36
Table 5: Training duration selected per final model. Selection epochs sum all validation fits. M uses three fits, so its total is not directly comparable to a single fit.
Model
5m
10m
15m
20m
25m
30m
60m
120m
C6
1.35
1.25
0.61
1.07
0.92
0.76
0.58
0.61
C12
1.21
0.96
0.91
1.08
0.96
0.60
0.69
0.17
C24
1.06
0.99
0.82
0.85
0.62
0.75
0.43
0.36
U6
1.05
0.92
0.77
0.88
0.84
0.54
0.63
0.61
U12
1.08
1.14
0.84
1.01
0.90
0.83
0.75
0.76
U24
1.48
0.91
0.75
0.90
0.83
0.85
0.87
0.57
Table 6: Mean observed-label residual Rank IC (%). Bold denotes the largest point estimate in each column, including the ensemble. C/U/R/M denote contiguous/uniform/recency/multi-recency, and the number denotes selected training months.
Figure 3: Observed-label Rank IC across all configurations and horizons.
Horizon
Best individual
Rank IC (%)
Ensemble IC (%)
5m
R6
1.90
1.53
10m
R6
1.57
1.45
15m
R6
1.00
1.10
20m
M12
1.26
1.31
25m
M24
1.08
1.17
30m
M24
1.09
0.98
Table 7: Retrospective best individual model versus the equal forecast ensemble.
Sampling
5m
10m
15m
20m
25m
30m
60m
120m
Contiguous
6
6
12
12
12
6
12
6
Uniform
24
12
12
12
12
24
24
12
Recency
6
6
6
12
12
12
24
24
Multi-recency
6
6
24
12
24
24
24
24
Table 8: Training-month quantity with the largest mean Rank IC within each sampling policy. Some differences are small, so these maxima should not be interpreted as significance tests.
Figure 4: Training quantity within each policy. Each panel retains its policy’s validation procedure while historical pool breadth changes with quantity for sampled methods.
Horizon
H1 2025
H2 2025
2026
All
5m
3.72
0.28
1.01
1.53
10m
2.43
1.44
0.88
1.45
15m
1.49
1.23
0.78
1.10
20m
1.72
1.18
1.14
1.31
25m
1.66
0.84
1.09
1.17
30m
1.65
0.50
0.89
0.98
Table 9: Equal-ensemble Rank IC (%) across subperiods. 2026 ends on 20 October.
Figure 5: Monthly Rank IC for C6, R6, M24 and the equal ensemble. Monthly variation is substantial.
Comparison
Horizon
Mean gain (pp)
95% interval (pp)
Months won
R6 vs C6
5m
0.579
[-0.011, 1.226]
13/22
R6 vs C6
10m
0.334
[-0.067, 0.731]
14/22
R6 vs C6
30m
0.008
[-0.413, 0.489]
10/22
M24 vs C24
30m
0.325
[0.028, 0.632]
14/22
M24 vs C24
120m
0.364
[-0.201, 1.083]
9/22
Table 10: Exploratory paired-month bootstrap comparisons. Gains and intervals are percentage points of IC, not returns. No multiple-testing adjustment is applied.
Horizon
Slope
Intercept (bp)
Forecast SD (bp)
Label SD (bp)
5m
0.206
-0.006
1.21
21.97
10m
0.255
0.004
1.47
29.68
15m
0.235
0.016
1.53
35.43
20m
0.236
0.007
1.83
40.12
25m
0.197
0.022
2.02
44.56
30m
0.162
0.042
2.07
48.24
Table 11: Pooled observed-label calibration of the equal ensemble. Slopes were estimated for diagnosis only.
Figure 6: Realized residual returns regressed on ensemble forecasts. Unit linear calibration would have slope one and intercept zero, while the estimated slopes are substantially below one.
Horizon
Between-model corr.
30m persistence
Sign flips (%)
5m
0.598
0.434
34.0
10m
0.613
0.479
32.2
15m
0.573
0.474
32.3
20m
0.604
0.527
31.6
25m
0.560
0.492
33.3
30m
0.542
0.457
34.6
Table 12: Forecast diversity and stability. Between-model correlation compares different models at the same origin. Persistence compares ensemble forecasts across adjacent 30-minute origins while sign flips report the proportion of matched forecasts that change sign.
Figure 7: Pairwise 30-minute forecast correlations, averaged within origins and then across months. Shared architecture and inputs do not produce identical forecasts.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Model
5m
10m
15m
20m
25m
30m
60m
120m
C6
1.22
1.20
0.58
0.89
0.75
0.55
0.27
0.38
C12
1.14
1.10
1.01
1.09
0.90
0.44
0.58
0.39
C24
0.93
1.07
0.93
0.83
0.60
0.66
0.32
0.40
U6
0.97
0.93
0.79
0.88
0.78
0.41
0.50
0.59
U12
0.87
1.05
0.78
0.90
0.73
0.60
0.61
0.70
U24
1.20
0.84
0.71
0.62
0.61
0.50
0.53
0.42
Appendix
Table 13: Full-period Pearson IC (%).
Model
5m
10m
15m
20m
25m
30m
60m
120m
C6
1.93
1.52
0.75
1.26
0.98
0.94
0.72
0.92
C12
1.86
1.58
1.31
1.30
1.16
0.93
0.75
0.28
C24
1.27
1.44
1.26
1.06
0.86
0.80
0.67
0.51
U6
1.65
1.15
0.89
1.08
1.00
0.68
0.49
0.62
U12
1.33
1.45
1.08
1.29
1.21
1.10
1.05
1.16
U24
1.61
1.01
0.73
0.80
0.71
0.78
0.74
0.62
Appendix
Table 14: 2025 Rank IC (%).
Model
5m
10m
15m
20m
25m
30m
60m
120m
C6
0.63
0.92
0.45
0.84
0.86
0.54
0.40
0.22
C12
0.41
0.18
0.42
0.80
0.72
0.19
0.63
0.04
C24
0.80
0.43
0.27
0.59
0.33
0.68
0.13
0.17
U6
0.30
0.65
0.63
0.63
0.64
0.36
0.80
0.60
U12
0.76
0.75
0.55
0.68
0.51
0.49
0.38
0.26
U24
1.33
0.78
0.76
1.03
0.98
0.94
1.05
0.50
Appendix
Table 15: 2026 Rank IC through 20 October (%).
Model
5m
10m
15m
20m
25m
30m
60m
120m
C6
3.33
2.02
0.91
1.26
1.13
1.38
1.03
1.46
C12
3.26
2.34
1.61
1.37
1.67
1.59
0.99
0.58
C24
3.34
1.99
1.31
1.34
1.34
1.28
0.99
0.85
U6
3.18
2.10
1.44
1.68
1.27
1.17
0.54
0.99
U12
3.55
1.88
1.11
1.41
1.53
1.68
1.34
1.58
U24
3.11
1.49
0.45
1.26
1.19
1.55
1.19
1.00
Appendix
Table 16: H1 2025 Rank IC (%).
Model
5m
10m
15m
20m
25m
30m
60m
120m
C6
0.58
1.05
0.59
1.25
0.83
0.52
0.42
0.42
C12
0.53
0.86
1.03
1.23
0.67
0.30
0.52
-0.01
C24
-0.70
0.91
1.21
0.78
0.39
0.35
0.35
0.19
U6
0.19
0.23
0.37
0.51
0.74
0.21
0.44
0.27
U12
-0.78
1.04
1.05
1.16
0.90
0.54
0.77
0.76
U24
0.18
0.55
1.01
0.36
0.26
0.04
0.30
0.27
Appendix
Table 17: H2 2025 Rank IC (%).
Portfolio
One-way / NAV (%)
Two-way / NAV (%)
Equal-sign
1,042.9
2,085.7
Prediction-weighted
1,332.6
2,665.3
Q5–Q1
1,466.7
2,933.5
Appendix
Table 18: Mean daily 30-minute target turnover of the equal ensemble.
Time series forecasting models are typically compared on pointwise error, which scores a prediction in isolation from the decision it is produced for, and a lower forecast error does not imply a better decision downstream. A parallel debate asks whether modern transformer architectures forecast better than recurrent and other lightweight models. We compare linear, fixed recurrent, transformer, and mixing based architectures against recurrent networks evolved by neuroevolutionary architecture search, evaluating each on forecast accuracy and on the net return of a daily long/short strategy. All models are fit on a pooled panel, one network trained across the whole universe. Across four mid-cap portfolios and three trading years, the evolved networks rank first on both forecast accuracy and net trading performance, while the second most accurate model loses money once positions are formed and costs are charged. The advantage tracks a horizon match, since rank IC for the evolved networks rises from a one-day to a ten-day scoring horizon while every model above 300 parameters declines. They are also the cheapest end to end: a CPU-only search of 16 minutes yields 66-weight networks that predict in 10.8~μs on a Raspberry Pi Zero, against transformer baselines of up to 817,153 parameters that require GPU training.
Jonathan Chang, Zimeng Lyu
Union County Magnet High School Scotch Plains, New Jersey, USA · Kean University Union, New Jersey, USA
Accurate prediction of equity returns remains a major challenge in computational finance due to the non-stationary, nonlinear, and low signal-to-noise ratio nature of financial time series. This paper proposes a hybrid two-stage architecture that combines a long short-term memory (LSTM) network with an XGBoost gradient-boosted regressor for multi-horizon stock return prediction across a diversified panel of 14 U.S. equities spanning six industry sectors. The LSTM component, comprising two stacked layers with 64 hidden units, processes 60-day sliding windows of five sequential market features to produce 64-dimensional temporal embeddings that encode learned sequential market dynamics. These embeddings are concatenated with 14 hand-crafted technical indicators to form a 78-dimensional hybrid feature vector, which is subsequently passed to an XGBoost regressor tuned via 3-fold cross-validation grid search. The framework is trained on a multi-stock pooled corpus using strict chronological splits and per-stock MinMaxScaling to prevent look-ahead bias, and evaluated across four prediction horizons of 30, 90, 252, and 365 trading days. Experimental results demonstrate that the hybrid model achieves a test RMSE of 0.0949 on the 30-day horizon, roughly one-third that of the standalone LSTM baseline, while marginally matching or surpassing the XGBoost-Only baseline across the majority of stocks. Directional accuracy rises with horizon length, reaching 97.6% at 365 days; we show, however, that this largely tracks the high base rate of positive long-horizon returns in the sample, and we therefore benchmark directional accuracy against a naive always-positive predictor and treat the above-base-rate gap at short horizons as the more informative signal. A composite investment scoring framework derived from multi-horizon predictions is further proposed to support portfolio ranking and decision support.
We ask whether the apparent forecasting skill of frozen time-series foundation models (TSFMs) on intraday index-futures returns is incremental skill. In a pre-registered evaluation on E-mini Nasdaq-100 (NQ) and E-mini S&P 500 (ES) futures, a common panel of five zero-shot models (Chronos-Bolt, Chronos-2, TimesFM 2.5, Moirai 2.0 and TiRex), plus separate experiments with Kronos, is compared with a no-change forecast at 5-120 minutes, using native prices, log returns, volatility-standardised returns and a volume clock as inputs. On a blind set opened once, 0 of 160 model-representation-horizon-market cells beat no-change, and a Hansen SPA test gives p = 1.000 on both markets; a later retrospective hold-out, scored as a descriptive check, gives 0 of 120. Input representations change forecast magnitude but not accuracy, and a transformation-aware ridge regression does as well. Several standard measures nevertheless suggest skill. Price-level correlation exceeds 0.9996, but no-change scores higher. Within-day RankIC across forecast origins rewards a trivial revert-to-mean rule (+0.35 at 30 minutes); we derive the expected within-day covariance under a random walk in closed form, and a sign-randomised null reproduces about 95% of the observed value. One model beats HAR for 5-minute volatility, but not the EWMA its input already contains. On strongly planted synthetic signals the models recover part of the signal at 5 minutes, and recovery deteriorates at 30 minutes. Benchmark-level success of TSFMs should therefore not be assumed to carry over to low-signal domains such as intraday returns.