Evolve on the Host, Predict on the Edge: Deploying Online Neuroevolutionary Architecture Search for Cross-sectional Stock Return Prediction

Jonathan Chang; Zimeng Lyu

cs.CEOct 7, 2026

Evolve on the Host, Predict on the Edge: Deploying Online Neuroevolutionary Architecture Search for Cross-sectional Stock Return Prediction

Authors: Jonathan Chang, Zimeng Lyu

Organizations: Union County Magnet High School · Department of Computer Science and Technology, Kean University

Abstract

Accurate forecasting models are usually large, expensive to update online, and fixed in architecture once trained. We apply ONE-NAS, an online neuroevolutionary architecture search that evolves a population of small recurrent networks as each window of data arrives, to daily cross-sectional stock return prediction, and pilot it on a host and endpoint pipeline: the host runs the search and ships each generation's champion genomes over TCP/IP to a Raspberry Pi 4B, which predicts online. On the Pi a single champion predicts a 50-stock window in 24.6~~ms and the ensemble of 40 island champions in 556~~ms, far inside the daily decision cycle. On four panels of US mid-cap equities over 2022--2024, reading the population as a rank-mean ensemble of island champions returns $+27.5\%$ net of realised transaction costs, against $+11.3$ to $+14.8\%$ for online LSTM, online GRU and monthly-retrained LSTM baselines and $+4.5\%$ for the single best genome used in prior ONE-NAS work.

Figures & tables

Figure 1: Deployment pipeline: ONE-NAS search on the host, online prediction on the Raspberry Pi endpoint.

Figure 2: On-device cost per pass of the two deployment options, on log axes; the line is the mean over ten runs and the band the min–max envelope. The island-champion ensemble runs all 40 island champions,

39\times

the weights of the single champion.

	2022	2023	2024	2022–24	$t^{c}$
ONE-NAS ( $N{=}60$ )	$+6.9$	$+13.7$	$+7.0$	$+30.1$ ${\scriptstyle\,\pm\,0.9}$	$-1.63$
ONE-NAS ( $N{=}40$ )	$+6.1$	$+12.4$	$+6.3$	$+27.5$ ${\scriptstyle\,\pm\,1.1}$	–
ONE-NAS ( $N{=}20$ )	$+5.1$	$+13.3$	$+6.3$	$+27.4$ ${\scriptstyle\,\pm\,1.1}$	$0.10$
Online LSTM	$+1.2$	$+4.5$	$+3.8$	$+11.3$ ${\scriptstyle\,\pm\,1.8}$	$7.32$
Online GRU	$+2.1$	$+3.1$	$+5.0$	$+11.7$ ${\scriptstyle\,\pm\,2.2}$	$6.59$
Periodic LSTM (mo.)	$+1.4$	$+3.9$	$+6.9$	$+14.8$ ${\scriptstyle\,\pm\,1.0}$	$6.66$

Table 1: Net return (%) by trade year, mean over

V1

–

V4

and seeds (10 for ONE-NAS; 48, 40 and 40 for the online LSTM, online GRU and periodic LSTM, Appendix A.3 ). The 2022–24 column is one run over the window, not the sum of the yearly cells.

Figure 3: Cumulative net return, 2022–2024 (books restarted at window start). Buy and hold follows the prior-study benchmark convention.

Fleet	Single	Ensemble	$\Delta$ net ( $t$ )	$\Delta$ Sharpe ( $t$ )
60 isl.	$+1.0$ / $0.09$	$+30.1$ / $0.93$	$+29.1$ $(18.3)$	$+0.84$ $(14.1)$
40 isl.	$+4.5$ / $0.21$	$+27.5$ / $0.87$	$+23.0$ $(18.2)$	$+0.66$ $(14.7)$
20 isl.	$+4.7$ / $0.25$	$+27.4$ / $0.87$	$+22.6$ $(14.9)$	$+0.62$ $(14.9)$
40 isl. a	$+8.9$ / $0.30$	$+36.7$ / $1.22$	$+27.8$ $(7.5)$	$+0.92$ $(6.6)$

Table 2: Prediction rule, within-run paired: single global-best genome vs. island-champion rank-mean ensemble from the same runs (Overlapping rule). Net % / Sharpe;

\Delta

= ensemble

-

single, paired

t

over seeds.

Arm	$1\times$	$2\times$	$3\times$	$5\times$	B/E a
ONE-NAS ( $N{=}60$ )	$+30.1$	$+27.2$	$+24.4$	$+18.8$	$11.6\times$
ONE-NAS ( $N{=}40$ )	$+27.5$	$+24.7$	$+21.8$	$+16.0$	$10.6\times$
ONE-NAS ( $N{=}20$ )	$+27.4$	$+24.4$	$+21.5$	$+15.6$	$10.3\times$
Periodic LSTM	$+14.8$	$+11.4$	$+8.0$	$+1.2$	$5.4\times$
Online GRU	$+11.7$	$+8.3$	$+4.9$	$-1.8$	$4.5\times$
Online LSTM	$+11.3$	$+8.1$	$+4.9$	$-1.5$	$4.5\times$

Table 3: Cost sensitivity: net return (%) on 2022–2024 as the realised per-name transaction cost is scaled by the stated multiple.

Figure 4: Performance vs. island count on the evaluation span (2022–2024, Overlapping rule; 10 seeds

\times

4 panels at every width). Left: pooled net return (mean

\pm

SE). Right: Sharpe, mean with

\pm 1

seed-SD band, and the worst seed. The 40-island headline was fixed on the tuning span (Table 5 , Appendix B.1 ) before this curve was measured.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Value
Islands $N$	40 (headline); 8 (reg.)
Elites / island	8
Offspring / island / generation	5
Mut. / intra- / inter-island crossover	0.7 / 0.2 / 0.1
Repopulation	every 50 generations
Fitness	validation MSE

Appendix

Table 4: ONE-NAS configuration used in Section 5 , in the notation of Sections 3.2 and 4.2 , selected on the 2016–2019 tuning span (Appendix B.1 ). “reg.”: the registered 8-island configuration; abbreviations in parentheses as in Tables and 11 .

Islands	$n$	rank IC	Net%	Sharpe	MDD%
8	20	$+0.0132$	$+26.0$	$0.92$	$11.6$
16	20	$+0.0149$	$+31.5$	$1.11$	$10.2$
20	20	$+0.0157$	$+36.0$	$1.24$	$10.4$
40	40	$+0.0169$	$+36.2$	$1.21$	$10.4$
50	40	$+0.0170$	$+36.4$	$1.21$	$10.4$
60	40	$+0.0168$	$+35.7$	$1.19$	$10.8$

Appendix

Table 5: Island count on the tuning span, 2016–2019, Overlapping rule, island-champion ensemble: the span on which the headline width was chosen. The evaluation-span curve is Figure 4 .

Seed	Gen.	Weights	Infer.	Energy	Power	Idle	MSE
		(end)	(ms)	(mJ)	(mW)	(mW)
42	300	143	22.8	56.9	2507	2103	1.046
43	300	145	23.5	58.5	2498	2126	1.026
44	300	138	23.6	58.3	2495	2102	1.071
45	300	145	25.3	62.9	2501	2125	1.011
46	300	155	24.2	59.4	2479	2122	1.025

Appendix

Table 6: On-device cost of the single champion (global-best genome), per seed. Each run streams 300 generations to the Raspberry Pi 4B, which evaluates the genome on that generation’s 50-stock test window. Times and energies are per pass, averaged over the run; energy is whole-board energy, and idle power is metered at the start of each run (Section 5.1 ).

Generations	Weights	Infer. (ms)	Energy (mJ)	Power (mW)
0–74	57	15.2	37.1	2453
75–149	111	23.8	58.7	2484
150–224	135	28.0	69.2	2495
225–299	152	31.5	77.8	2490
Growth	2.7 $\times$	2.1 $\times$	2.1 $\times$	1.02 $\times$

Appendix

Table 7: Single-champion cost against search progress, pooled over the ten runs of Table 6 . Each row averages one quarter of the campaign. Inference time and energy grow with the evolved network; board power is flat, so the energy growth is time, not draw (Section 5.1 ).

Seed	Gen.	Weights	Infer.	Energy	MSE	MSE
			(ms)	(mJ)	(ens.)	(memb.)
45	268	4,528	527	1,589 (371)	0.975	1.018
46	300	4,317	444	1,378 (371)	0.975	1.011
47	300	4,584	618	1,796 (411)	0.975	1.019
48	300	4,521	634	1,845 (223)	0.975	1.020
49	300	4,651	602	1,746 (213)	0.975	1.016

Appendix

Table 8: On-device cost of the island-champion ensemble, per seed. A pass runs all 40 island champions and averages their predictions, so weights and costs are totals over those networks; energy is whole-board energy, with the marginal cost above idle in brackets. MSE (ens.) is the ensemble’s error and MSE (memb.) the mean error of its members. Seeds that stopped before generation 300 are averaged over the generations they completed (Section 5.1 ).

Option	Runs	Weights	Infer.	Energy	MSE
			(ms)	(mJ)
Single champion	10	114	24.6	61 (7)	1.031
Island-champion ens.	6	4,492	556.2	1,652 (328)	0.975
Ratio	–	39 $\times$	23 $\times$	27 $\times$ (45 $\times$ )	0.946 $\times$

Appendix

Table 9: Cost of one pass for the two deployment options, pooled over all generations of all runs; weights are averaged over the campaign (the per-quarter values of Table 7 average to 114). Energy is whole-board energy, with the marginal cost above idle in brackets (Section 5.1 ).

Fleet	$N$	$\rho$	Member $m$	Ensemble	Predicted
20 islands	20	0.166	$+0.0068$	$+0.0145$	$+0.0149$
40 islands	40	0.150	$+0.0065$	$+0.0153$	$+0.0156$
60 islands	60	0.163	$+0.0067$	$+0.0157$	$+0.0160$

Appendix

Table 10: Ensemble diversity mechanism (Section 5.3 ): member rank IC

m

, pairwise member prediction correlation

\rho

, measured ensemble rank IC, and the equicorrelated averaging prediction

m\sqrt{N/(1+(N-1)\rho)}

, 2022–2024, Overlapping rule.

Configuration	$\Delta$ net	Configuration	$\Delta$ net
C0 (registered)	$+3.9$	NTS 600	$+12.8$
Select: IC-gated	$+0.5$	Rounds 2	$+11.0$
Select: IC	$+2.8$	PER $\alpha$ 0.4	$+4.0$
Target: 5-day CS	$+4.5$	PER $\alpha$ 0.8	$+11.7$
5-day CS + IC-gated	$+1.8$	PER $\lambda{\times}3$	$+20.6$
Elite 5 / gen 10	$+9.6$	BP epochs 20	$+12.5$

Appendix

Table 11: Ensembling gain (ensemble

-

single champion, within-run, net % on 2016–2019, Overlapping rule) inside every trained hyperparameter configuration of the tuning screen; positive in 16/16 cells (Section 5.3 ).

Rule	2022	2023	2024	2022–24
Overlapping	$+6.1$ ${\scriptstyle\,\pm\,0.8}$	$+12.4$ ${\scriptstyle\,\pm\,0.7}$	$+6.3$ ${\scriptstyle\,\pm\,0.5}$	$+27.5$ ${\scriptstyle\,\pm\,1.1}$
Banded a	$+12.5$ ${\scriptstyle\,\pm\,1.2}$	$+14.6$ ${\scriptstyle\,\pm\,1.3}$	$+3.4$ ${\scriptstyle\,\pm\,0.9}$	$+33.9$ ${\scriptstyle\,\pm\,1.8}$
Cond. L/S b	$+1.3$ ${\scriptstyle\,\pm\,1.6}$	$+12.8$ ${\scriptstyle\,\pm\,1.0}$	$+10.8$ ${\scriptstyle\,\pm\,1.2}$	$+25.0$ ${\scriptstyle\,\pm\,1.8}$

Appendix

Table 12: Trading rule: net return (%) of the ONE-NAS ensemble (40 islands) under the three rules on the same predictions and costs, mean over

V1

–

V4

, 10 seeds; the Overlapping row is the 40-island row of Table 1 (Section 5.4 ).

Arm	bps/day	%/yr	NW $t$	Beta
ONE-NAS ( $N{=}60$ )	$+4.43$	$+11.2$	$+2.73$	$+0.120$
ONE-NAS ( $N{=}40$ )	$+4.05$	$+10.2$	$+2.56$	$+0.109$
ONE-NAS ( $N{=}20$ )	$+3.99$	$+10.0$	$+2.77$	$+0.114$
Periodic LSTM (monthly)	$+2.39$	$+6.0$	$+2.01$	$+0.046$
Online LSTM	$+1.93$	$+4.9$	$+1.61$	$+0.069$
Online GRU	$+1.75$	$+4.4$	$+1.91$	$+0.033$

Appendix

Table 13: Factor-adjusted alphas: pooled daily book returns of each arm of Table 1 regressed on the Fama–French three factors plus momentum, 2022–2024, Overlapping rule, Newey–West

t

with lag 10 (Section 5.4 ).

Explore similar work

Forecast Accuracy Is Not Trading Profit: Evolving Small Recurrent Networks for Stock Return Prediction
Oct 6, 2026Jonathan Chang, Zimeng LyuForecasting Backbone
Union County Magnet High School Scotch Plains, New Jersey, USA · Kean University Union, New Jersey, USA
A Hybrid LSTM-XGBoost Framework for Multi-Horizon Stock Return Prediction Across Diversified Equity Portfolios
Sep 14, 2026Seif ElDein Mostafa, Yahia Ahmed, Farah Datwish +1Cnn-Lstm Xgboost Model
Faculty of Computer Science MSA University Giza, Egypt
From Regression to Inference: Meta-Learning Predictors for Neural Architecture Search
May 10, 2026Liping Deng, MingQing XiaoNeural Architecture Search Meta-Learning
Department of Mathematics University of California, Riverside Riverside, CA 92507 USA · School of Mathematical and Statistical Sciences Southern Illinois University Carbondale Carbondale, IL 62901 USA

Oct 6, 2026, cs.NEJ/K move · Enter open · S save

Forecast Accuracy Is Not Trading Profit: Evolving Small Recurrent Networks for Stock Return Prediction

Jonathan Chang, Zimeng Lyu

Union County Magnet High School Scotch Plains, New Jersey, USA · Kean University Union, New Jersey, USA

Forecasting Backbone

Open paper AskPDF arXiv

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~

μ

s on a Raspberry Pi Zero, against transformer baselines of up to 817,153 parameters that require GPU training.

Abstract

Explore similar work

Forecast Accuracy Is Not Trading Profit: Evolving Small Recurrent Networks for Stock Return Prediction

A Hybrid LSTM-XGBoost Framework for Multi-Horizon Stock Return Prediction Across Diversified Equity Portfolios

From Regression to Inference: Meta-Learning Predictors for Neural Architecture Search