Variant Perception

Where We Disagree With the Market

The market is paying $92.85 per ADS for a Halyk Bank multiple on a sell-side narrative that says "deep discount to NU/MELI peers will close" — and the disagreement is not whether the discount should close but whether consensus actually believes it will. The headline $126.33 Benzinga consensus price target is arithmetically dragged up by a stale May-2024 $175 print from New Street; the three most-recent ratings (Freedom Broker, JPM, JPM) average ~$93 — the same as spot. What looks like a 36% upside-to-consensus is in fact a market that has already quietly converged with the bear: most-recent calls are HOLD/Neutral at $87-$95, and the post-Tencent/post-IG-bond rally has merely closed the gap to where current consensus already sat. The real consensus belief is "fairly priced for Kazakhstan single-country EM fintech, awaiting Section 404(b) and Hepsiburada validation" — and the most-decision-relevant gap in the file is that the report's NPL coverage evidence (99% → 90% → 80% over three years while gross NPLs grew 91%) is mathematically inconsistent with the stable 2.2% Cost-of-Risk reading consensus models accept at face value.

The resolution path is dated and narrow: the 1H 2026 IFRS interim disclosure (August-September 2026) and the first independent Section 404(b) auditor attestation (March 2027 20-F) settle the load-bearing variant view. Between them, the Hepsiburada engagement print and the Rabobank A.Ş. closing window (mid-2026) test the second-largest disagreement. No variant view in this file survives the bull-versus-bear flatness frame Stan settled on — instead, this page identifies where consensus itself is mispriced, not where bull beats bear.

Variant Perception Scorecard

Variant strength is moderate — high on the NPL/CoR mathematical gap, moderate on the consensus-mirage and Hepsiburada-as-impairment claims, lower on the structural-discount-versus-dislocation point because the IG bond demand and Tencent stake have already partially repriced country risk. Consensus clarity is mixed: the public consensus number ($126.33 average) and the actual consensus behavior (most-recent ratings near spot, no upgrades, two downgrades from JPM and Susquehanna in the last 9 months) point in opposite directions, which is the disagreement itself. Evidence strength is highest on the NPL math and the goodwill exposure, both of which are direct line-items in the FY2025 20-F. The first resolving event is the 1H 2026 IFRS interim in August/September 2026 — four months out from the May 27, 2026 date of this file.

Consensus Map

The market belief on KSPI is not monolithic; the public "consensus PT $126" frame and the actual analyst behavior in the last 12 months differ. The table below maps the implied underwriting assumptions embedded in the price tape, the sell-side average, and management's own guidance switch.

The consensus map shows where market belief is clear and disagrees with evidence (rows 2, 3) versus where market belief is itself muddled (rows 1, 4, 6). The single sharpest disagreement is row 2 — the CoR-stability reading rests on a model whose ICFR failed once and whose NPL coverage trajectory contradicts the stability claim arithmetically.

The Disagreement Ledger

Three ranked disagreements survive the materiality, evidence, resolution-path, and disconfirming-signal tests. A fourth was considered (regulatory ratchet on BNPL pricing via ARDFM) but was dropped — the rule's form and effective date are undecided, so it is currently a watch item, not a variant claim.

Disagreement #1 — The CoR/NPL coverage math is broken

Consensus would say: "Cost of Risk has been stable at 2.0-2.2% across the cycle, Q1-26 printed 0.7%, the data-flywheel underwrites in under six seconds at industry-leading loss rates, and management explicitly attributes the coverage decline to a portfolio mix shift toward lower-coverage BNPL products." The arithmetic disagrees on a single point: provisioning growth (₸243B → ₸372B, +53% over two years) has lagged NPL growth (₸244B → ₸467B, +91%) — that is what coverage falling from 99% to 80% means. A model that under-builds reserves into a growing NPL stock produces a flattering CoR by construction. If the variant view is right, the FY2026 provision line carries an ~₸89B catch-up that compresses pretax margin by 8-9 percentage points, and the data-flywheel claim rests on an ECL model whose internal controls have failed in production once. The cleanest disconfirming signal is the 1H 2026 NPL coverage print: above 85% with CoR holding under 2.5% refutes the variant; at or below 80% with CoR drifting confirms it.

Disagreement #2 — Sell-side consensus is a stale-average mirage

Consensus would say: "Six covering analysts average $126.33, implying meaningful upside to spot — the stock is trading well inside the bull case." The arithmetic disagrees: that average is anchored by a $175 New Street print from May 7, 2024 — four months before Culper, ten months before Hepsiburada consolidated, and twelve months before the Krivenok class action was dismissed. Strip the pre-Culper print and the average collapses toward the most-recent three ratings ($87, $88, $96 — mean $90.33), which is where the stock already trades. Stock Analysis lists 7 analysts at $100.75; AlphaSpread 1y-average is $102.61; Public.com shows 2 analysts at HOLD with a $91.50 target. Susquehanna downgraded the stock from Positive to Neutral in November 2025 (PT $130 → $87) and only raised PT to $95 in April 2026 while keeping Neutral. JPM cut PT $96 → $88 in late 2025. If we are right, the "deep discount to consensus" framing in the bull case is a coordination problem — the live coverage already prices the stock fairly, and the upside-from-here trade requires an upgrade cycle, not a discount-closure. The cleanest disconfirming signal is a coverage-list price-target raise above $110 from a top-tier broker (JPM, Susquehanna, Morgan Stanley) on the back of the August Q2 print.

Disagreement #3 — Hepsiburada is consensus-framed as growth, evidence-framed as impairment risk

Consensus would say: "Q4-25 order growth +19% and FY2025 Adj-EBITDA approaching breakeven validate the Türkiye transplant; Tencent's April 2026 secondary purchase ratifies the super-app playbook; management's FY2026 Adj-EBITDA breakeven target is conservative." The evidence points to three under-weighted structural facts. First, HEPS engagement is 6.7 purchases per consumer per year against Kazakhstan's 24.8 — one-quarter the intensity after a full year of integration. Second, ₸429.7B of HEPS goodwill (17% of equity) sits against a 1.4% EBITDA margin and a ₸89.6B segment net loss in a market where Trendyol is roughly 3× scale and Alibaba-funded. Third, Rabobank A.Ş. — the deposit-funded fintech leg that is the strategic reason for the entire Türkiye build-out — has slipped once and has no firm new closing date. If the variant view is right, the HEPS goodwill is one bad operating quarter from a formal impairment test, the Türkiye thesis is a 3-7-year proof-of-concept rather than a 2026 inflection, and Adj-EBITDA breakeven against a low single-digit Turkish-lira denominator (flattered by IAS 29 monetary gains) is not the same as proving the moat transplants. The cleanest disconfirming signal is the BDDK closing of Rabobank A.Ş. inside the mid-2026 window with the $300M capital injection executing on plan.

Evidence That Changes the Odds

The table below pulls the highest-leverage evidence items in the file and shows how a consensus reader and a variant reader interpret each. Each item is sourced to a specific upstream tab or data point; the fragility column names the way the evidence could mislead.

How This Gets Resolved

Six signals — each observable in a dated disclosure, regulatory filing, sell-side action, or analyst-tracked metric — close the variant view in one direction or the other. Two are leading (Rabobank closing and sell-side estimate revisions through August); two are decisive (1H 2026 NPL coverage and Section 404(b) attestation); two are confirmatory (HEPS engagement and goodwill test). The 1H 2026 NPL coverage print is the highest-decision-weight resolver because it tests the single largest arithmetic gap in the file on the shortest timeline.

What Would Make Us Wrong

The NPL coverage variant view depends on the mix-shift framing being substantially false. It might not be. Kaspi's product mix has genuinely tilted toward shorter-duration BNPL — a category that does have a structurally lower coverage requirement than traditional auto or unsecured cash loans because the loss-given-default is lower (small ticket, fast amortisation). If management discloses product-by-product coverage data on the 1H 2026 call and shows that within each cohort coverage is stable, then the headline coverage decline is a denominator effect of growth-mix and the variant view collapses. The Q1-26 CoR of 0.7% is consistent with that read. The bear's best receipts (FY23 ICFR weakness in this exact area; allowance growth lagging NPL growth on the consolidated number) survive the test only if the disclosed cohort data shows coverage falling within at least one product line.

The "consensus mirage" variant view depends on no new bullish coverage initiation in the August window. A buy-side initiation from a top-tier broker at a $130-150 PT post-Q2 print would re-anchor the average meaningfully — coverage lists this small (6-7 analysts) are unstable, and a single high-conviction Buy with high target rebuilds the discount-closure narrative. The Tencent strategic stake is the most plausible catalyst for such an initiation (asian-coverage broker picks up the name on the back of the WeChat-heritage connection). If that lands, the "actual consensus is HOLD at spot" framing dies inside a single broker-research note.

The Hepsiburada-as-impairment variant view depends on Türkiye's order trajectory continuing to decelerate or the Rabobank close slipping again. Both are unknowable today. If Q2-26 HEPS prints sustained +20%+ order growth, engagement crosses 8+ purchases per consumer, and BDDK approves Rabobank inside the mid-2026 window with the $300M injection executing on plan, the impairment-risk frame loses analytical bite — the table grows into the goodwill at the rate the bull case assumes. The variant survives only if the engagement gap is structural (Trendyol's lead is too entrenched) rather than transitional.

The honest red-team summary: variant #1 (NPL math) has the strongest evidence; variant #2 (consensus mirage) has the most observable resolution; variant #3 (Hepsi impairment) has the most distant and noisy resolution path. A PM should weight conviction in that order — and accept that all three could be refuted by a clean 1H 2026 print and a clean Section 404(b) attestation, in which case the bull case re-emerges and the multiple has room to migrate from 7.8× toward the 12-15× range. The asymmetry is real but the burden of proof is on the variant view, not on the market.

The first thing to watch is the 1H 2026 NPL coverage disclosure in the August-September 2026 IFRS interim 6-K — coverage above 85% with product-by-product mix detail closes the largest arithmetic gap in the file; coverage at or below 80% with CoR drifting above 2.5% validates the highest-conviction variant view in this report.