Stralgo — Case Study · Salih Devran Tırık

01Context

What it is, why it matters, who it’s for.

What

A no-code platform where traders compose strategies on a visual canvas and replay years of market data with bit-exact reproducibility.

Why

Backtests are only useful if they replay identically. Drift breaks trust and quietly invalidates research.

For

Quant-curious traders without a Python background, and analysts who need to share runs that reviewers can reproduce.

02Problem

Samestrategy,samedata,samehash.Everytime.

Constraints

01Single 8-core node — no horizontal scaling for v1

02Sub-second feedback while editing 100k-node graphs

03Versioned, forkable strategies with instant rollback

04Non-engineering users on the editing surface

Difficulty

Determinism in trading sims is a quiet, expensive problem. Floats, RNGs, system clocks, even map iteration order will drift a result enough to invalidate a year of research.

03Approach

Decisions, in order of stakes.

Fix every clock, every float

Replaced `time.Now()` with a monotonic event clock, swapped float64 for fixed-point Q32.32, and seeded a deterministic RNG keyed by run hash.

Strategies as immutable hashes

Every strategy compiles to a canonical JSON body and is stored content-addressed. Forks reference parents; the engine only ever runs by hash.

Engine in Go, UI on the side

Replaced the Python prototype with a zero-allocation Go engine. The SvelteKit canvas talks to it through a thin RPC layer and never blocks on a sim.

04Execution

The sim engine is a single tight loop. Indicators update in place against pre-sized ring buffers, the strategy emits a typed signal, and the orderbook settles fills in fixed-point arithmetic.

Engine core — zero-alloc stepgo

// engine/sim.go
type Sim struct {
    clock  Tick           // monotonic, integer microseconds
    book   *Orderbook     // fixed-point Q32.32
    rng    *DeterministicRand
}

func (s *Sim) Step(b Bar) Result {
    s.clock = b.At
    for _, ind := range s.indicators {
        ind.Update(b)        // zero alloc, pre-sized ring buffers
    }
    sig := s.strategy.Eval(s.indicators)
    if sig.Kind != Hold {
        s.book.Submit(sig.ToOrder(s.clock))
    }
    return s.book.Settle(b)  // exact fills, no float drift
}

Strategies are stored as hashes. A run never references a mutable row; reviewers can re-execute by hash months later and get the same equity curve down to the satoshi.

Content-addressed strategiessql

-- strategies (immutable, content-addressed)
create table strategies (
  hash       bytea primary key,         -- sha256 of canonical json
  parent     bytea references strategies(hash),
  body       jsonb not null,            -- compiled DSL
  author_id  uuid  not null,
  created_at timestamptz default now()
);

create index on strategies (parent);
-- forks are cheap; runs reference the hash, never the row id.

Pipeline · canvas → DSL → engine → report

05Artefacts

Visual rule composer with live indicator preview.

Equity, drawdown and trades in one report.

Branch and fork strategies like source control.

07Outcomes

Result

12×faster,bit-exact,andquietlyforkable.

Impact

Closed beta with 80 traders. Every reported run is reproducible by hash. P95 API stayed under 100ms through the launch month while strategies grew to 2.3k versions.

Backtest speed12×vs. Python baseline5-year tick dataset, single core

Determinism100%bit-exactIdentical SHA256 across 1k re-runs

P95 API84ms−71%Strategy save / fetch under load

Active strategies2.3k+Versioned across closed beta

08Takeaways

Learned

Determinism is a feature, not a tax. Lock it in early or rebuild later.

Hashing strategies turned versioning, forking, and audit into the same primitive.

Fixed-point is unfashionable and saves more arguments than it loses.

Would change

→

Ship a visual diff between strategy versions earlier — users want to see what changed before re-running.

→

Move the analyzer registry behind a signed plugin API in v1, not v2.

Continue

Next case

SeoTrace

Lead Engineer · Full-stack · 2025 · SaaS · SEO tooling

← Freelance Works All works ↗