Why I Started an AI Quant Research Project in 2023

In 2023 I organised a research effort around an AI-driven quantitative analysis and automated trading system. I did not do it because I believe algorithms abolish uncertainty. I did it because the bottleneck in modern markets is no longer only idea scarcity—it is the cost of processing messy, high-frequency, multi-source information without drowning in noise.

Manual analysis still matters. Thesis formation, institutional knowledge, and ethical judgment do not emerge from a loss function. But manual analysis is slow at tasks machines handle well: scanning large feature sets, detecting regime shifts in data streams, stress-checking thousands of scenarios, and enforcing rules without fatigue. The project’s premise is complementary labour, not human replacement theatre.

The technical stack we care about is familiar on paper—machine learning, deep learning where it earns its keep, big-data pipelines, and enough compute to iterate honestly. The less glamorous parts matter more: data provenance, leakage control, walk-forward discipline, and a refusal to confuse backtest beauty with deployable edge. An impressive chart from a contaminated experiment is not research; it is self-persuasion.

Automation in trading also needs a philosophy of control. I am interested in systems that can propose, size, and execute within hard risk envelopes—not systems that “feel” their way into unbounded leverage. Intelligent risk management, dynamic position sizing, and automated execution only deserve those adjectives if the constraints are explicit and testable. Otherwise automation merely accelerates mistakes.

Sentiment and alternative data are tempting because they feel modern. They are also littered with spurious correlations. Our bias is to treat such signals as hypotheses that must survive out-of-sample scrutiny and economic plausibility checks. If a model cannot explain, even roughly, what economic behaviour it is capturing, I am reluctant to let it touch real capital.

Timelines matter for honesty. The work is still in development. A commercial launch window in the second half of 2027 is a plan, not a promise carved in stone. It depends on technical validation, risk testing, and regulatory compliance. Anyone who has shipped financial software knows those gates are real. I would rather under-promise on calendar romance than over-claim readiness.

There is a cultural reason for the project as well. Across Asia—and in Indonesia in particular—I see rising interest in markets alongside uneven access to professional tooling and education. Building research capacity is not only about a single product. It is about creating a place where methods can be taught, challenged, and improved. Tools without literacy become toys or traps.

I reject two extremes. One extreme says AI will print money if you merely subscribe. The other says AI is irrelevant because markets are efficient enough already. Reality sits awkwardly in the middle: markets are competitive, edges decay, and better information processing can still matter at the margin if paired with risk controls and humility. That margin is worth researching precisely because it is not magical.

What success looks like for me is not a marketing screenshot of a perfect equity curve. It is a system that makes analysis more consistent, surfaces risks earlier, and leaves a clear audit trail of why a decision was made. If it also improves returns after costs, good. If it only improves decision hygiene, that alone can be valuable.

Artificial intelligence will not make investing easy. Used well, it may make certain kinds of negligence harder to sustain. That is reason enough to build.

Talent mix matters as much as models. A team that only ships code without market scar tissue will optimise the wrong objective. A team that only debates markets without engineering discipline will never industrialise learning. I want both languages in the same room, with shared vocabulary for risk.

Evaluation design is where many AI finance projects quietly fail. If you peek at test data while tuning, you will congratulate yourself on luck. We budget time for locked evaluations and for adversarial reviews that try to kill a signal before capital does.

Execution research is not an afterthought. A signal that cannot survive realistic slippage, partial fills, and venue outages is a blog post, not a strategy. Automation must include degradation modes: reduce size, widen checks, halt.

I also care about explainability for stakeholders. Not every model needs full interpretability, but every deployed decision path needs a story a risk officer can challenge. Black boxes that cannot be interrogated do not belong near client or personal capital without hard wrappers.

The 2027 target window keeps us honest about sequencing: research, validation, compliance, then wider use. Sequence discipline is itself a form of risk management.