Alkame Inc.
Calibrated, conformal-prediction market intelligence across the NIFTY 500 — an applied research program DBERT incubates in exchange for equity.
At a Glance
₹1.2Cr
Research contribution — DBERT services delivered in exchange for equity
4 Months
May – August 2026 applied research program
6,300+ Stocks
NIFTY 500 universe, sub-100ms ingestion latency
01
Why Calibrated Confidence Matters
A documented gap between unregulated retail-facing signals and validated confidence — grounded in SEBI’s own regulatory data.
+41%
~91%
~20%
~2%
Sources: SEBI, Jul 2025; CFA Institute, 2025.
Alkame-Nifty50 tests one falsifiable question: does enforcing calibrated confidence as a display gate — via split conformal prediction — and walk-forward, cost-aware validation as a promotion gate, produce lower calibration error and comparable-or-better risk-adjusted edge than Microsoft’s Qlib and a naive rule-based baseline?
This is a pre-registered study design. No empirical results are claimed here — the research team has committed in advance to report the outcome whichever direction it points.
02
Where This Sits vs. Existing Work
Alkame-Nifty50’s contribution is architectural and applied, not a new algorithm — it integrates findings that, on their own, stop short of a production system.
Kaya & Nguyen (2025)
Conformal prediction achieves error rates tracking nominal significance levels, on 764 US large/mid-cap stocks, Jan 2022–Jan 2024.
Gap this project addresses: Small-scale, US-only proof of concept — not India-specific, not integrated into a production pipeline with drift detection.
Bailey et al. (2014)
Formal proof that backtest overfitting is a near-inevitable statistical artifact of testing many configurations.
Gap this project addresses: Diagnoses the problem rigorously, but is a methodology paper — no live architecture for continuously re-validating deployed models.
Gama et al. (2014)
Establishes the standard taxonomy of concept drift (sudden, gradual, incremental, recurring) in ML broadly.
Gap this project addresses: General-purpose survey, not applied to retail financial signal generation or tied to audit/regulatory requirements.
03
How Alkame-Nifty50 Is Built
Four layers, each built to resist a specific, documented failure mode in retail-facing quantitative trading tools.
Model Layer
Per-symbol, per-horizon ensembles — logistic regression, random forest, gradient-boosted trees, and a stacking combiner — trained under strict no-lookahead discipline.
Calibration Layer
Split conformal prediction sits on top of every model output. A prediction is only surfaced once its prediction set collapses to a single label — a distribution-free coverage guarantee, not a raw softmax score.
Drift Detection & Retraining
Every prediction — shown or withheld — is logged with a resolvable outcome. A retrained model must re-clear the same validation gates before replacing production, never auto-promoted.
Validation Gate
Walk-forward backtesting across multiple non-overlapping windows, with explicit transaction-cost and slippage modeling, gates any promotion.
04
Technical Difficulties & How We Approached Them
Five specific, well-documented failure modes shaped this design — each paired with the concrete choice made to address it, and the research it draws on.
Problem — Most retail-facing prediction tools report a point prediction with no validated measure of confidence.
Approach — Split conformal prediction as a display gate — a call is only shown once its prediction set collapses to a single label, chosen over Platt scaling or isotonic regression for its distribution-free coverage guarantee.
Kaya & Nguyen, 2025
Problem — A single attractive backtest is a well-documented statistical trap — testing many configurations can make a strategy look strong purely by chance.
Approach — Every candidate model clears a walk-forward, multi-window, cost-and-slippage-aware validation gate before promotion. One good backtest is necessary, never sufficient.
Bailey, Borwein, López de Prado & Zhu, 2014
Problem — Live markets drift, so a model validated last quarter can silently decay without anyone noticing.
Approach — A Prediction Ledger logs every prediction against its eventual resolved outcome. Sustained degradation triggers retraining — the retrained model re-clears the same gates, no shortcuts.
Gama et al., 2014
Problem — Comparing a new architecture fairly against existing tooling, rather than a strawman.
Approach — The evaluation protocol benchmarks against Qlib — Microsoft’s open-source quant research platform — on identical data and time windows, plus a naive rule-based floor.
Yang, Liu, Zhou, Bian & Liu, 2020
Problem — India’s regulatory footing for algorithmic and "black box" signal providers changed materially in 2025.
Approach — Scope is deliberately restricted to signal-quality assessment, not automated order execution — staying on the lower-friction research/signal side of SEBI’s new principal-agent framework.
SEBI, Feb 2025
05
Built Against a Moving Regulatory Target
SEBI’s February 2025 circular introduced India’s first governance framework for retail algo trading — this changes how a system like Alkame must be classified.
Under SEBI Circular SEBI/HO/MIRSD/MIRSD-PoD/P/CIR/2025/0000013 (Feb 2025), stockbrokers are the principal, and any algo provider acts as their agent. Algorithms are classified as disclosed or undisclosed (“black box”)— Alkame’s ML ensemble falls in the latter category by construction.
The project stays on the research/signal side of this line rather than automated execution — the compliance layer hard-blocks any output that reads as personalized advice or unauthorized order placement, avoiding the higher-friction execution regulatory bracket.
SEBI, Feb 2025
06
A 4-Month Applied Research Program
May through August 2026, moving from infrastructure and model-building through live validation and drift monitoring.
MAY 2026
Data infrastructure and feature-store setup across NIFTY 500 constituents; no-lookahead discipline established.
JUNE 2026
Model ensembles and the conformal calibration layer built out; Qlib and naive baselines prepared for matched comparison.
JULY 2026
Walk-forward validation protocol executed; Prediction Ledger begins logging live, resolvable outcomes.
AUGUST 2026
Live drift-monitoring window and program wrap-up; findings reported per the pre-registration commitments, in either direction.
Interns Currently Working on Alkame
Meet the DBERT interns actively contributing to this research program.
Loading contributors…
Want to Work on Projects Like This?
DBERT interns get hands-on placement on real, funded research programs like Alkame-Nifty50 — not simulated coursework. Apply through the internship platform to get matched to a live project.