{"id":"26a62381-2acd-49b3-8eff-501e03ad5b68","arxiv_id":"2412.05297","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A machine-learning DSS using fundamental ratios plus macro features achieves 70% average test accuracy in predicting whether a Tehran Stock Exchange stock will beat fixed-income returns, and uses this signal for stock-bond-gold allocation.","lead":"The paper describes a decision support system for stock selection and asset allocation on the Tehran Stock Exchange, combining company financial ratios, trading statistics, and Iranian macroeconomic indicators. It reports higher classification accuracy than three earlier fundamental-analysis methods and shows positive asset allocation results during 2022-2024.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline comparison substitutes generic NN for cited methods; claimed superiority over prior research is unsupported.","rationale":"The reader's weakest assumption correctly identifies the baseline representation as the most load-bearing concern. The paper's central claim is that the proposed DSS is superior to novel researches (contribution 3). To support this, the evaluation must compare against the actual methods from the cited papers. Instead, Section 5.5 states that baselines are compared by 'using their features and training a model for stock prediction,' which means the original model architectures, loss functions, hyperparameters, and problem formulations are discarded. This conflates feature-set comparison with method comparison. The reported accuracy gap and portfolio performance could be driven by the choice of a generic NN and the binary threshold formulation rather than by the proposed method's intrinsic merit. This concern is directly settled by re-running the original baseline models on the same TSE data. I considered other issues such as missing error bars, ad hoc strategy parameters, and potential data snooping, but those are secondary: even if error bars were added, the comparative claim would remain invalid if the baselines are not faithfully represented. Therefore, the reader's conditional verdict is appropriate, and no verdict change is needed.","tokens_in":12428,"tokens_out":5435,"duration_ms":47995,"concrete_test":"Re-run the exact baseline models from Tsai et al. (regression with 18 financial ratios predicting next-quarter return), Abe et al. (deep learning with 25 features predicting one-month cross-sectional return), and Arkan (ratio-based trend prediction) on the same TSE data, same train/test split, and same evaluation metric (accuracy of beating the fixed-income ETF return, plus the top-20 portfolio experiment). If the proposed method still outperforms all three original methods by a margin comparable to Table 6, the superiority claim survives; otherwise, the reported gains are an artifact of the baseline implementation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.5 compares against 'Baseline-1' (Tsai et al.), 'Baseline-2' (Abe et al.), and 'Baseline-3' (Arkan) by 'using their features and training a model for stock prediction.' The original methodologies are never reproduced: Tsai's regression model, Abe's deep network with a one-month horizon, and Arkan's specific procedure are all replaced by a generic neural network (one hidden layer, 100 neurons, binary cross-entropy) trained on each paper's feature list. Thus the comparison measures the predictive content of feature sets under the authors' own model class, not the performance of the cited methods. The central claim of superiority to prior research (contribution 3) is therefore not established by the reported results: the higher test accuracy (70% vs 58/62/60) and the portfolio return superiority could stem from the model architecture, the binary classification threshold against fixed-income ETFs, or other modeling choices, rather than any genuine advantage over the actual baselines. This is load-bearing because it directly undermines the paper's comparative claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a decision support system (DSS) for stock selection and asset allocation on the Tehran Stock Exchange. The system collects, cleans, and integrates fundamental reports, stock trading features, and macroeconomic indicators, then trains a one-hidden-layer neural network to predict whether a stock will beat the fixed-income ETF return over horizons of 1 to 12 months. The authors compare this model against feature sets from three cited fundamental-analysis studies and against five classical classifiers, and they further use the stock-level predictions to construct a market-direction probability and two asset-allocation strategies (70/20/10 and 10/70/20 across stocks, gold, and bonds). The central claims are that the proposed method has higher test accuracy than the baselines (70% vs 58%, 62%, 50% in Table 6) and that the asset-allocation strategies outperform the market index, gold, and baseline portfolios in the test period.","tokens_in":12620,"tokens_out":5338,"duration_ms":48210,"significance":"If the comparative results were properly established, the paper would make a useful applied contribution for high-inflation, medium-sized equity markets: the end-to-end data pipeline is substantial, the feature set is broad and domain-relevant, the problem formulation against a fixed-income hurdle is sensible for investors, and the evaluation in both nominal and real terms is appropriate. The authors also correctly incorporate the one-month reporting lag, which strengthens the simulation's realism. However, the evidence currently does not support the headline superiority claim because the baselines are not implemented as the original methods, and no uncertainty quantification is given. The asset-allocation module also lacks out-of-sample validation of its handpicked threshold and weights. The underlying system is promising, but the paper needs major strengthening before its central claims can be accepted.","major_comments":[{"comment":"The comparison does not implement the cited baseline methods. The text states that baselines were formed by 'using their features and training a model for stock prediction,' meaning each cited feature set is fed to the authors' own one-hidden-layer ANN. Tsai et al.'s regression model, Abe et al.'s deep-learning regression with a one-month horizon, and Arkan's specific procedure are not reproduced. Therefore Tables 5 and 6 show the predictive content of feature sets under a common model class, not the superiority of the proposed method over prior approaches. The abstract and conclusion claim 'superiority to novel researches'; this claim must either be re-stated as a feature-set comparison or supported by running the original methods on the same TSE data.","section":"§5.5, Tables 5–6"},{"comment":"No uncertainty quantification is reported. All accuracies come from a single chronological split, with no confidence intervals, bootstrap, repeated seeds, or significance tests. At the 1-month horizon the proposed method (64%) is actually below Baseline-2 (70%), and several differences in Table 5 are in the 1–4 percentage-point range. The headline test-accuracy gap of 70% vs 58–62% could be within sampling noise. Please report means and standard deviations over multiple runs and a paired significance test (e.g., McNemar) on the same test instances.","section":"§5.5, Table 5"},{"comment":"The label construction is under-specified. Formula (1) defines the monthly fixed-income return, but the classification horizon varies from 1 to 12 months (Table 5). The paper does not state how the fixed-income hurdle is compounded over the prediction horizon or how overlapping quarterly report dates are handled for multi-month labels. This is load-bearing because the class labels are defined by comparing stock returns with this hurdle. Please give the exact label rule for every horizon.","section":"§5.2, Eq. (1) and §5.3"},{"comment":"The asset-allocation module has no out-of-sample validation of its decision rule. The 50% probability threshold and the 70/20/10 and 10/70/20 weights are presented as fixed choices; no procedure for selecting them or sensitivity analysis is given. The test period is a single 27-month window, and Figures 10–11 show cumulative results without transaction costs, fees, or rebalancing frictions. The claim that the strategy outperforms gold and market indices therefore needs robustness checks (e.g., varying the threshold and weights, adding costs), or it should be presented as an illustrative simulation rather than as a validated result.","section":"§6.1–6.2, Table 7, Figs. 10–11"},{"comment":"The market-direction predictor is evaluated only visually. Figure 7 plots the TSE index against predicted probabilities, and the text says 'the model performance is good', but no quantitative accuracy, hit rate, or comparison to a random or naive benchmark is given for the aggregate market prediction. Because the asset-allocation strategies in §6.2 are triggered by this prediction, this performance measure should be made explicit and statistically evaluated.","section":"§6.1, Figure 7"}],"minor_comments":[{"comment":"The sentence 'the adopted model is a an artificial neural network' contains a typo and should be rephrased.","section":"§5.3"},{"comment":"The Beta formula is given as Cov(Rm,Rs)/Var(Rs); the standard definition is Cov(Rm,Rs)/Var(Rm). If the feature was computed with the printed formula, this should be corrected or clarified.","section":"Table 1"},{"comment":"It is unclear whether the columns of Table 5 are averages over all test symbols and periods or macro-averages; please define how the accuracy is aggregated.","section":"§5.5, Table 5"},{"comment":"The legends refer to 'Proposed method' for both the stock-selection portfolio and the asset-allocation strategy; please clarify which series corresponds to which strategy.","section":"§6.2, Figures 10–11"},{"comment":"The 'Stock Market Exchange' feature appears in the feature list, but all data come from the Tehran Stock Exchange; explain why this feature is not constant or remove it.","section":"§5.1, Table 2"},{"comment":"The paper does not provide a data-availability statement, code, or a reproducibility appendix; given the specialized data pipeline, this would substantially increase the value of the submission.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"This is an engineering-oriented DSS paper with a promising system but a currently unsupported comparative claim. The most important fix is to either faithfully reimplement the baselines or explicitly reframe the contribution as a feature-set comparison under a common model. Adding uncertainty quantification and robustness checks for the asset-allocation rule are also necessary. The manuscript would benefit from careful copyediting and a clearer description of the label construction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Take the comparison with prior work with a grain of salt. The paper's own out-of-sample numbers on Tehran Stock Exchange are plausible, but Section 5.5 compares against feature sets from Tsai, Abe, and Arkan, not their actual methods—every baseline is reimplemented as a generic one-hidden-layer ANN. So the claimed 'superiority to novel researches' is really 'our features work better than their features under our model.' That's a meaningful result, but it's not the same claim.\n\nWhat's genuinely new: an end-to-end DSS that handles the messy crawling/cleaning of Iranian financial statements, a feature set mixing ratios, market trading data, and macro variables, and a target defined as beating fixed-income ETF return. That last choice is a good fit for high-inflation markets and is a sensible problem formulation for fundamental investors. The paper gives a clear architecture and the empirical evaluation on TSE 2015-2024 is a new dataset result.\n\nThe soft spots beyond baselines: no confidence intervals or significance tests, so the 70% vs 62% test-accuracy gap could be noise. The allocation weights (70/20/10, 10/70/20) appear handpicked to fit the test period; no sensitivity analysis. No code or data. The market-direction plot (Fig 7) is a single backtested window with no uncertainty quantification. Some references are sloppy (e.g., 'Malkiel' cited for random walk).\n\nNone of this is fatal to the core direction. The stress-test's concern about baseline substitution is correct and it does directly undercut contribution 3. The reader's conditional verdict is fair.\n\nFor peer review: I'd send it, but with a required revision: either run the original baseline methods on the TSE data or explicitly re-frame the comparison as feature-set ablation. Also demand error bars. If they do that, this could be a useful paper for practitioners in small/medium economies. As is, I wouldn't cite the superiority claim, but I might point to the problem formulation.","headline":"A workmanlike applied DSS paper whose comparative claim overstates what the experiments show; the out-of-sample TSE result is worth a second look.","tokens_in":13138,"tokens_out":3434,"would_cite":false,"duration_ms":29924,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A decision support system that models fundamental financial-report data as a classification against fixed-income ETF returns reaches 70% test accuracy on the Tehran Stock Exchange and generates stock/bond/gold allocation strategies with…","keywords":["Stock Prediction","Fundamental Analysis","Decision Support Systems","Asset Allocation","Financial Markets","Tehran Stock Exchange","Financial Ratios","High Inflation Economies"],"falsifier":"Run the original methods behind the three cited baseline studies—their actual regressions, network architectures, and training procedures, not just their feature lists—on the same 413-stock Tehran Stock Exchange dataset and forecast horizon, and count their average test accuracy and cumulative portfolio return. If any of those original models matches or exceeds the proposed method's 70% test accuracy or its cumulative portfolio return, the paper's central claim of superiority to prior research fails.","tokens_in":12237,"feed_emoji":"📈","tokens_out":6418,"duration_ms":57929,"temperature":0.7,"pith_summary":"The paper tries to establish that an end-to-end pipeline for collecting, cleaning, and modeling fundamental financial-report data can predict which stocks on the Tehran Stock Exchange will outperform a fixed-income ETF over horizons of one to twelve months, and that this prediction can drive a simple asset-allocation rule across stock, bond, and gold ETFs. The proposed neural-network classifier, trained on financial ratios, stock-type indicators, trading activity, and macroeconomic variables, reaches 70% average test accuracy, compared with 58%, 62%, and 50% for three baseline feature sets drawn from prior fundamental-analysis studies when each is fed to the same generic network. In a twenty-stock portfolio experiment over the test period, the method's cumulative return is reported to exceed those of the baselines and of the stock market, equal-weight index, and gold fund. A sympathetic reader would care because the work addresses a practical investor need: mid- to long-term stock selection in a high-inflation emerging market, using the kind of semi-structured report data that many prediction studies avoid.","feed_headline":"Stock picker using financial reports hits 70% accuracy","feed_subtitle":"A decision-support pipeline turns ratios and macro data into stock/bond/gold strategies that outpaced the index and gold.","key_machinery":"The load-bearing object is the classification target itself: label a stock as 1 if its return beats the fixed-income ETF over the forecast horizon, with the ETF's monthly outside-option return computed as $(1+\\mathrm{YTM})^{1/12}-1$. This target encodes an investor's alternative—a low-risk bond-like return—so the model learns relative attractiveness rather than absolute price. Around it sits a feature set of 23 financial ratios, including gross profit growth and revenue growth, plus stock-type indicators, trading-activity measures, and macroeconomic variables such as the USD/IRR exchange rate, gold return, government bond return, and market index return. The market-direction signal aggregates per-stock probabilities weighted by market capitalization, and the allocation rule thresholds that signal at 50%.","core_discovery":"On its own terms, the discovery is that a classification problem—will a stock's return over the next one to twelve months exceed the return of a fixed-income ETF, computed from yield to maturity as $(1+\\mathrm{YTM})^{1/12}-1$—is a more learnable and more decision-relevant formulation than the raw-return regression used in prior fundamental-ratio studies. Using 413 Tehran Stock Exchange stocks from 2015 to 2024, with report publication lagged one month to simulate real-world timing, the authors train a one-hidden-layer neural network with 100 neurons on 23 financial ratios plus stock-type, trading, and macroeconomic features. They report 70% average test accuracy against 58%, 62%, and 50% for the baselines, and better cumulative returns for a portfolio that periodically selects the twenty highest-probability stocks. They also derive a market-level probability by market-cap weighting the per-stock probabilities, and use that signal to switch between a 70/10/20 stock/bond/gold allocation and a 10/70/20 allocation.","pith_inferences":["The paper does not run the original baseline models on the Tehran data—it feeds each prior paper's listed features into the same generic neural network—so the reported superiority should be read as a claim about the feature set and classification formulation, not about the full pipeline versus the original methods.","A direct ablation the paper does not report—removing the macroeconomic feature category, or removing the trading-activity category, and re-measuring test accuracy—would quantify which part of the 70% edge comes from macro variables such as USD/IRR and gold returns versus from financial ratios alone.","The asset-allocation results come from a single test window in which global gold prices were rising, and the strategy always keeps 20% in gold; the real-return advantage over gold may not survive a stable or falling gold regime.","The classification target against fixed-income returns likely transfers to other high-inflation and emerging markets, but its advantage may shrink in low-inflation markets where the bond outside option is less central to stock selection."],"forward_implications":["On the Tehran Stock Exchange test period, the model's 70% test accuracy means a simple rule—buy stocks whose predicted beat-the-ETF probability exceeds 50%, otherwise hold fixed-income ETFs—would have selected the better asset class in most quarters.","The market-cap-weighted market probability can act as a leading indicator: the paper's Figure 7 shows the predicted probability dropping to single digits in several quarters of market weakness.","The twenty-stock portfolio built from the model's top predictions realized higher cumulative nominal and real returns than the aggregate market index, the equal-weight index, or the gold fund over the 2022–2024 test window.","The asset-allocation strategies produce positive real returns under high inflation because capital shifts to fixed-income and gold when the predicted market probability falls below 50%, and back to stocks when it rises above 50%.","Because the target embeds the fixed-income outside option, the same architecture could be applied directly to any market where low-risk bond or deposit returns define the investor's benchmark, not just to high-inflation economies."],"supporting_citations":[{"why":"Supplies Baseline-1: 18 financial ratios used for quarterly stock-return prediction, whose feature set the proposed method is compared with.","marker":"[45]"},{"why":"Supplies Baseline-2: a deep-learning cross-sectional return-forecasting method with 25 financial-ratio features, the second comparison point.","marker":"[1]"},{"why":"Supplies Baseline-3: a financial-ratio approach to predicting price trends in an emerging market, the third comparison feature set.","marker":"[3]"}],"fun_headline_variants":["Fundamental ratio model hits 70% accuracy on Tehran stocks","Neural net classifies stocks vs bond yield with 70% hit rate","New DSS picks stocks and switches assets using fundamentals","Stock classification beats bond return in Tehran backtest","Financial statement analysis yields 70% accuracy in stock pick"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the comparison setup—feeding each prior paper's listed features into the same generic neural network—fairly represents those papers' true predictive power; if the original methods were rerun on the same Tehran Stock Exchange data, the reported gap could shrink or reverse.","fun_headline_variants_meta":{"raw":{"variants":["Fundamental ratio model hits 70% accuracy on Tehran stocks","Neural net classifies stocks vs bond yield with 70% hit rate","New DSS picks stocks and switches assets using fundamentals","Stock classification beats bond return in Tehran backtest","Financial statement analysis yields 70% accuracy in stock pick"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000509,"raw_usage":{"total_tokens":2523,"prompt_tokens":1037,"completion_tokens":1486,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":1404}},"tokens_in":653,"tokens_out":1486,"duration_ms":12445,"temperature":1.0,"reasoning_tokens":1404,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:45:46.193391+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the original methods behind the three cited baseline studies—their actual regressions, network architectures, and training procedures, not just their feature lists—on the same 413-stock Tehran Stock Exchange dataset and forecast horizon, and count their average test accuracy and cumulative portfolio return. If any of those original models matches or exceeds the proposed method's 70% test accuracy or its cumulative portfolio return, the paper's central claim of superiority to prior research fails.","supporting_citations":[{"cited_title":"Stock selection using machine learning based on financial ratios","cited_arxiv_id":null,"evidence_quote":"Supplies Baseline-1: 18 financial ratios used for quarterly stock-return prediction, whose feature set the proposed method is compared with."},{"cited_title":"Deep learning for forecasting stock returns in the cross-section","cited_arxiv_id":null,"evidence_quote":"Supplies Baseline-2: a deep-learning cross-sectional return-forecasting method with 25 financial-ratio features, the second comparison point."},{"cited_title":"The importance of financial ratios in predicting stock price trends: A case study in emerging markets","cited_arxiv_id":null,"evidence_quote":"Supplies Baseline-3: a financial-ratio approach to predicting price trends in an emerging market, the third comparison feature set."}],"review_version":1}