A risk desk could use the calibration result, with care. In the 250-day rolling ablation, LASSO cuts a quantile network's inputs from 48 candidates to a handful, and its Kupiec record rises from 0 of 6 passes to 4 of 6.
The title raises climate risk in ESG portfolios. The climate diagnostics cover IHSG only.
A risk model without a trade
Syalsabila, Ikhwana and Rahmanda compare SRI-KEHATI, the longest-running of the exchange's ESG indices with 25 screened issuers, against the broad IHSG and its 900+ listed companies. They ask whether the indices carry climate tail risk differently. Their proposed mechanism is mediation: a flood or fire shock could reach index returns through domestic equity prices first, leaving a variable selector with no separate climate signal even if climate matters. The model produces daily VaR at the 1%, 5% and 10% quantiles. It has no trading rule.
LASSO screens 50 candidate predictors, including eight climate series: physical and transition risk indices, daily geopolitical risk (GPR), ERA5 temperature anomaly, CHIRPS rainfall, EUA carbon, natural gas and clean-energy equity returns. A quantile regression neural network (QRNN) then estimates VaR. This small net uses pinball loss and has one output per quantile; it retrains daily on the prior 250 observations. KernelSHAP assigns each day's estimate across its inputs. Listwise deletion leaves 1,499 days from September 2014 to December 2025 out of 3,146 raw. Training takes 1,199 observations; the 300-observation test runs from 8 November 2023 to 23 December 2025.
LASSO selects no climate variable at any quantile for either index. For SRI-KEHATI, it retains LQ45 and JK Finance returns, adding ISSI at 5% and 10%. For IHSG, it retains those three plus the CDS spread and JK Mining at 1%; Nikkei joins at 5%, and Hang Seng and gold at 10%. When three climate series are forced into the IHSG network, their SHAP shares stay near flat at 13.08%, 13.32% and 12.73% across the quantiles. In a separate linear quantile regression using all eight, the share climbs from 6.12% at 10% to 17.25% at 1%.
Where did the climate signal go?
The authors state the limit in their abstract: whether SRI-KEHATI screening isolates climate risk "was not tested; mediation covers IHSG only." On the ESG index, their climate evidence goes no further than LASSO variable selection.
For IHSG, the raw correlation between clean-energy returns and the index is 0.121; after six controls it is -0.0005. LQ45 and JK Finance, two of those controls, are large-cap sub-indices of IHSG itself. The authors acknowledge that index composition may explain the collapse. They also note that "clean-energy returns are themselves equity returns, so co-movement with domestic equity factors is expected."
I read the collapse as mostly mechanical.
The forced-in climate share raises another problem. It averages 7.0% over the 28 days of the El Niño 2023 window and 7.9% during the 2024 floods. In the residual period, which has no named event, it reaches 16.4%. The authors say they "have no explanation for this pattern."
Selection remains hard to pin down
The network architecture is specified: Linear(16) and Linear(8) layers, batch norm, ReLU, 0.1 dropout, a target z-scored within each window, and an output bias warm-started at that window's empirical quantile. It has roughly 323 parameters with SRI-KEHATI's 6 inputs and 403 with IHSG's 11. Four implementation questions remain:
- Selection loss. The paper describes LASSO selection per quantile, yet retains a squared-loss LASSO-OLS after three penalized quantile variants went non-sparse. Its CV curve shows pinball loss, with selected strengths of 1.3219 (SRI-KEHATI) and 0.1417 (IHSG) at the 5% quantile. We could not work out how the squared-loss fit produces three different variable sets.
- Gaps. Listwise deletion removes 52.4% of the sample, leaving non-contiguous days. The 300 test observations stretch over about 26 months, or roughly 515 to 535 trading days; lag and realized-vol terms therefore skip dates. The authors flag the issue and did not test sensitivity.
- Training settings. Seed, epochs, learning rate and batch size are said to match across runs. We did not find their values in the text. Code is available only on request.
- Which panel. The main model starts in September 2014. The ablation starts in January 2013 and excludes Bitcoin and aluminum. The panels differ, and their results disagree.
We could not run the models ourselves. SRI-KEHATI, IHSG and the Indonesia Stock Exchange (IDX) sub-indices used by every selected model are outside our data. US proxies would test a different mechanism.
Same-day inputs and the pinball wins
Across all 12 Diebold-Mariano comparisons in the main backtest, the network beats Historical Simulation, the empirical quantile of the trailing 250-day window, and GARCH(1,1)-t. The statistics range from 3.44 to 8.17, all p<0.001. The authors give two qualifications. The QRNN uses same-day closes unavailable to the benchmarks, and the ablation fails to reproduce the edge. In that run, 11 of 12 comparisons are insignificant. Historical Simulation wins for IHSG at 10% (DM -2.51, p=0.012).
The authors say the comparison "supports the pipeline's fit, not a like-for-like forecasting advantage." I agree. Fit is cheap when an IHSG 1% VaR uses today's LQ45 close, which becomes known only after the loss.
Their ablation also shows the weakness of pinball loss as a scoreboard, as the authors acknowledge. Raw-QRNN feeds the same network all 48 candidates without LASSO and gets the lower loss, 0.000903 against 0.001096. Yet it breaches the 10% SRI-KEHATI quantile 20.88% of the time. The authors attribute that to under-dispersed, insufficiently negative VaR estimates. The benchmarks are refitted monthly; the network retrains daily.
What survives on SRI-KEHATI
On 249 common dates, LASSO-QRNN breaches SRI-KEHATI at 0.80%, 5.22% and 10.44%. The corresponding Kupiec p values are 0.747, 0.874 and 0.817. IHSG passes only at 10% (12.85%, p=0.149); it fails at 1% (3.61%) and 5% (10.04%). At 5%, the direction changes between runs. The main model under-violates at 2.0% (6 of 300), while the ablation over-violates at 10.04%. The authors have not traced the discrepancy. They concede that pre-selection "does not by itself deliver adequate calibration" for IHSG, and they compared no other regularizer.
Those 249 common dates yield 2 to 32 violations per cell. A replication using lagged sub-index returns, with SRI-KEHATI still passing Kupiec at three levels, would change my view of what the paper has shown.