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

1 Introduction

• Accurate online forecasting models are often large, making them expensive to update and serve, and their architectures remain fixed after training.The passage cites Das et al. 2023, Ansari et al. 2024, and Woo et al. 2024 for large accurate forecasting models.Source: §1 · p. 1
• Stock forecasting faces low signal-to-noise returns and regime shifts, while scheduled refits and incremental weight updates leave the architecture fixed in advance.The cited work on shifting feature-to-return relationships is Zhao, Kong, and Shen 2023.Source: §1 · p. 1
• ONE-NAS evolves recurrent-network populations as data windows arrive, with offspring briefly trained from inherited weights and survivors selected on recent data.The paper contrasts this with a single best genome, which trails neural baselines, and a rank-mean island-champion ensemble, which outperforms them.Source: §1 · p. 1
• The paper applies ONE-NAS to daily cross-sectional stock return forecasting on CRSP data, with a host running search and a Raspberry Pi endpoint running online inference.The contributions include a deployment workflow for forecasting and trading on forecasts.Source: §1 · p. 1

2 Related Work

• Online forecasting and incremental-learning methods adapt weights or mixture weights while keeping architectures fixed before the stream begins.Examples include FSNet (Pham et al. 2022), OneNet, Proceed (Zhao and Shen 2025), and incremental learning (Zhao, Kong, and Shen 2023).Source: §2 · pp. 1–2
• ONE-NAS moves evolutionary architecture search online, while prior work read its population through a single best genome; this paper combines online search with ensemble reading.The cited ensemble-search work found architecturally diverse ensembles more robust under dataset shift, but had not combined that idea with online architecture search.Source: §2 · p. 2
• Hardware-aware NAS and once-for-all training finish search before deployment, while on-device learning updates weights of a fixed network.The passage distinguishes the paper's host-run continuous search and generation-wise shipment of small networks from these approaches.Source: §2 · p. 2
• The paper identifies a deployment gap: stock prediction models are usually evaluated by backtest, with edge-hardware serving measurements rarely reported.It reports a Raspberry Pi pilot as part of its host-and-endpoint pipeline.Source: §2 · p. 2

3 Method

• ONE-NAS maintains a population of RNNs whose architectures and weights evolve as observations arrive, and this work applies it to daily stock-return prediction across a host and endpoint.Its genome graph grows from a minimal input-to-output seed, with recurrent-cell nodes and feedforward or recurrent edges.Source: §3 · p. 2, §3.1 · p. 2

3.1 Online Neuroevolution

• Genomes grow through node and edge mutations or crossover; retained components inherit weights, and independently evolving islands preserve population diversity.Brief offspring training is enabled by weight inheritance; islands exchange genetic material only through low-rate inter-island crossover.Source: §3.1 · p. 2

3.2 Online Protocol

• Each window triggers a generation: offspring train on earlier history, genomes are scored on recent validation data, and selected genomes predict the next window before its labels are revealed.The history split slides forward after the window closes, and the passage states that no genome is trained on data it is scored on or selected on data it later predicts.Source: §3.2 · p. 3

3.3 Prediction Rules

• The single-champion rule uses the global best genome, whereas the ensemble averages the daily ranks of each island's best elite because output scales are not comparable.The prediction rules are a scoring-time choice on the same search runs.Source: §3.3 · p. 3

3.4 Deployment Architecture

• The host trains and selects genomes, then transmits them over TCP/IP; the Raspberry Pi predicts online while the host trains the next generation.The endpoint does no training and can keep predicting with the last received genomes if the host is unavailable.Source: Figure 1 · p. 2, §3.4 · p. 3

4.1 Dataset

• Four disjoint portfolios each contain 50 randomly drawn S&P 500 stocks from 203 eligible constituents with market capitalisation of 2–50 billion USD and at least 20 years of history.The CRSP panels run from early 2004 through the end of 2024.Source: §4.1 · p. 3
• Models use seven daily per-stock inputs and predict next-day returns rank-normalised across stocks, matching trading rules that act on relative ordering.Inputs include daily return and rank, bid–ask spread, illiquidity, 20-day prior return, turnover acceleration, and 21-day volatility.Source: §4.1 · p. 3

4.2 ONE-NAS Setup

• ONE-NAS uses five-day windows and one generation per window; offspring train for 10 epochs on 40-day subsequences and fitness is validation-block MSE.The setup uses 40 islands and six recurrent cell types; burn-in spans 2004–2015, tuning 2016–2019, and evaluation 2020–2024, with reported results on 2022–2024.Source: §4.2 · pp. 3–4

4.3 Baselines

• The four baselines are online LSTM and GRU models, a monthly retrained LSTM, and an online first-order autoregressive model, using the same inputs, target, trading rules, and costs.Each baseline is a single model at the scale of ONE-NAS's evolved networks.Source: §4.3 · p. 4

4.4 Trading Rules and Metrics

• Three trading rules use only daily predicted cross-sectional rankings, and reported returns subtract per-stock, per-day CRSP bid–ask-spread transaction costs.Metrics include rank IC, net return, annualised Sharpe ratio, turnover, alpha, and market beta.Source: §4.4 · p. 4

5 Results

• ONE-NAS hyperparameters were selected on the disjoint 2016–2019 tuning span; results generally use 40 islands across four panels and ten seeds on 2022–2024.The default evaluation uses the Overlapping rule and the island-champion ensemble.Source: §5 · p. 4

5.2 Net Return against Baselines

• +27.5% net return for the 40-island ensemble exceeded the neural baselines’ +11.3% to +14.8%; its paired margin was +12.7 to +16.2 points.The ensemble led every baseline in 2022 and 2023; in 2024, the monthly-retrained LSTM drew level at +6.9% versus +6.3%.Source: Table 1 · p. 5, §5.2 · p. 5

5.1 On-Device Inference Cost

• 24.6 ms for a single champion and 556 ms for the 40-island ensemble per 50-stock pass kept both options inside the daily decision cycle.The ensemble uses 39× the weights and 23× the time; its MSE of 0.975 beats the single champion’s 1.031.Source: §5.1 · pp. 4–5, Figure 2 · p. 4

5.3 Prediction Rule: Single Champion versus Island-Champion Ensemble

• +27.5% net return for the ensemble versus +4.5% for the single champion shows that reading the evolved population as an ensemble changes performance substantially.The ensemble’s Sharpe was 0.87 versus 0.21, a paired gain of +23.0 points (t = 18.2); the gain was positive in all 16 tuning-screen configurations.Source: Table 2 · p. 5, §5.3 · p. 5

5.4 Trading Rule, Transaction Costs and Factor Exposure

• The ensemble remained profitable through 10–12× realistic costs, while online LSTM and GRU returns fell below zero at 5×.Banded returned +33.9% and Conditional long–short +25.0%, versus +27.5% for Overlapping; Overlapping remained the rule used throughout. Factor-adjusted ensemble alphas were +10.0% to +11.2% per year at market betas of 0.11–0.12.Source: §5.4 · pp. 5–6, Table 3 · p. 6

5.5 Operating the Pilot

• +26.5% for a continuously evolving population versus +12.3% for a 2022 cold start indicates the value of pre-evaluation evolution in this configuration.Returns and Sharpe rose steeply from 10 to about 20 islands, then flattened; widths of 20–60 were not separable at ten seeds, while per-seed Sharpe dispersion fell as width increased.Source: §5.5 · p. 6, Figure 4 · p. 6

6 Conclusion

• ONE-NAS evolves the population on a search host and sends each generation’s champions to a Raspberry Pi, which forecasts online.The pipeline is proposed for stock forecasting and trading on the resulting forecasts.Source: §6 · p. 6
• 24.6 ms for a single champion and 556 ms for 40 island champions to predict a 50-stock window on the Pi, within the daily cadence.The authors identify adding an execution layer, measuring realised costs live, and obtaining industry user feedback as future work.Source: §6 · p. 6
• +27.5% net return for the island-champion ensemble versus +4.5% for the single champion and +11.3% to +14.8% for neural baselines on four US mid-cap panels over 2022–2024.The paired gain over the single champion was +23.0 points (t = 18.2); the ensemble outperformed the baselines under the same rule and costs (t > 6.5).Source: §6 · p. 6

A Experimental Details

• Appendix A describes the panels, features, target, ONE-NAS configuration, baselines, trading rules, metrics, and paired tests.Its subsections correspond to Sections 4.1 and 4.2 and the trading-rule and metric descriptions.Source: A Experimental Details · p. 9
• Appendix B reports the hyperparameter selection on the tuning span and tables supporting Sections 5.1, 5.3, and 5.4.Source: A Experimental Details · p. 9
• The main text carries every number it relies on, while the appendix documents how those numbers were produced.Source: A Experimental Details · p. 9

A.1 Dataset Construction

• From 203 eligible S&P 500 constituents, 200 were randomly drawn and split evenly into four nonoverlapping portfolios; differing historical start dates affect burn-in only, with common tuning and evaluation calendar spans.Each portfolio has 50 constituents, and all four end on 2024-12-31.Source: §A.1 · p. 9
• Seven causal daily features describe returns, liquidity, trading activity, and volatility; the next-day target is return rank-normalised across stocks.Cross-sectional ranking addresses drifting feature levels and matches trading rules that use only stock ordering.Source: §A.1 · p. 9

A.2 ONE-NAS Configuration and Hyperparameter Search

• Each island retains eight elites and produces five offspring per generation; offspring receive ten training epochs on prioritized replay windows of length 40.Every 50 generations, the island with the worst best-elite fitness is repopulated from the global best genome.Source: §A.2 · pp. 9–10

A.3 Baselines

• The baseline set includes online LSTM and GRU, monthly-retrained LSTM, and online autoregressive models; reported baseline figures average independent runs.The baselines use the same panels, prequential protocol, features, target, trading rules, and transaction costs as ONE-NAS.Source: §A.3 · p. 10

A.4 Trading Rules

• All trading rules act on each day’s predicted stock ordering, making reported metrics invariant to strictly monotone output transforms.Overlapping is used throughout Section 5; Banded and Conditional long–short are also run on the same predictions in Section 5.4.Source: §A.4 · p. 10
• Overlapping buys the top 10 and shorts the bottom 10 daily, holding each portfolio for 10 trading days; Banded holds top-10 entrants until rank falls past 35.Positions are netted across overlapping portfolios before trades, and costs are charged on changes in aggregate per-name notional. Conditional long–short’s sign gate is always met on rank-normalised predictions, reducing it to daily full rebalancing.Source: §A.4 · pp. 10–11
• Transaction costs use each stock’s time-varying half bid–ask spread, converted to a cost per dollar traded and charged on traded notional.All reported returns are net of these costs; the cost sensitivity analysis scales the per-name costs for every stock and day.Source: §A.4 · p. 11

A.5 Metrics and Statistical Comparison

• Rank IC is the mean daily Spearman correlation between predicted and realised cross-sectional returns; net return is cumulative profit and loss after transaction costs as a fraction of initial capital.Annualised Sharpe scales mean daily return per unit of its standard deviation by 252; turnover is mean daily traded notional relative to gross book size.Source: §A.5 · p. 11
• Maximum drawdown is the largest peak-to-trough decline in cumulative net return; the tuning-span island-count table covers 2016–2019 under the Overlapping rule.The table caption identifies the evaluation-span curve as Figure 4 but provides no cell values.Source: §A.5 · p. 11, Table 5 · p. 11
• Factor exposure is assessed by regressing pooled daily book returns on Fama–French three factors and momentum; the intercept is alpha and the market coefficient is beta.Standard errors use Newey–West with lag 10 to account for autocorrelation induced by the 10-day holding period.Source: §A.5 · p. 11
• The comparison unit is a (panel, seed) cell, with paired t-tests run over shared cells and the shared-cell count stated when seed sets differ.Unless stated otherwise, a ONE-NAS configuration has 40 cells from 10 seeds on four panels.Source: §A.5 · p. 11

• Appendix B.1 reports the tuning-span configuration screen, while B.2–B.4 provide tables behind the on-device, ensembling, and trading-rule results.Source: B Additional Results · p. 11

B.1 Hyperparameter Selection on the Tuning Span

• The six tested widths were 8, 16, 20, 40, 50, and 60 islands; comparisons on the five seeds shared across widths preserved the observed ordering.Source: §B.1 · p. 12
• From 20 to 60 islands, rank IC spans +0.0168 to +0.0170 and net return spans +35.7 to +36.4%, with widths not separable.Sharpe ranges from 1.19 to 1.24 across those widths, and maximum drawdown stays between 10.2 and 11.6% across every width.Source: §B.1 · p. 12
• The 8-island setting is lower than 40 islands by 0.0037 in rank IC and 10.2 points of net return.Source: §B.1 · p. 12
• Forty islands was chosen on the tuning-span plateau; 50 and 60 were indistinguishable there and each adds an endpoint network.The evaluation-span curve was measured after the configuration choice and did not inform it.Source: §B.1 · p. 12
• Prioritised replay was retained because uniform sampling showed no tuning-objective advantage (∆IC −0.0024, t = −1.9); no other variant met the paired-improvement bar.Source: §B.1 · p. 12

B.2 On-Device Measurements

• On Raspberry Pi 4B, one 50-stock-window prediction takes the single champion 24.6 ms (MSE 1.031); averaging 40 island champions yields ensemble MSE 0.975, lower than every member’s.The ensemble costs 23× the time and 27× the whole-board energy per pass; marginal-energy ratios are omitted because campaigns used different power supplies.Source: §B.2 · p. 12, Table 8 · p. 13

B.3 Source of the Ensembling Gain

• Across widths, the equicorrelated-averaging prediction is within 3% of measured ensemble IC; ensembling is positive in 16 of 16 tuning configurations, averaging +8.3 points.Those tuning-screen ensembles had 8 members; the 40-island tuning-span fleet’s gain was +27.8 points.Source: §B.3 · p. 12

B.4 Trading Rule and Factor Exposure

• Table 12 reports net returns for the 40-island ensemble under three trading rules, while Table 13 reports factor-adjusted alphas for each Table 1 arm.The factor regression uses Fama–French three factors plus momentum, for 2022–2024 under the Overlapping rule.Source: §B.4 · p. 13, Table 12 · p. 13, Table 13 · p. 13