// notes on tech

Blog

Survivorship bias: the failures leave the dataset 2016 today ACME in your data BETA in your data CORP ✗ delisted — gone from the file DELT in your data ECHO ✗ bankrupt FOXX ✗ acquired at a loss a backtest run today picks only from the survivors — and never takes their losses
Quant

Survivorship bias — backtesting on a graveyard

Your historical data contains the companies that made it. The ones that went to zero quietly left the dataset — and took the losses your strategy would have taken with them.

2026-08-09 · 6 min read
Look-ahead bias: a decision reaching into data that did not exist yet not knowable yet knowable at decision time now you decide here ← the close your code read it hasn't happened yet
Quant

Look-ahead bias — the one-bar mistake

The most flattering bug in quant finance: your strategy quietly uses information it couldn't have had yet. It hides in a single line, and it always looks like a great year.

2026-08-08 · 6 min read
Queue position at a price level resting bids at 100.00 · first in, first out incoming sell 500 400 all filled 300 100 filled you · 100 no fill the tape prints 500 @ 100.00 your backtest: "it touched — filled!" 700 shares were ahead of you; only 500 traded
Quant

Did that fill even happen?

Your backtest fills your limit order because the price touched it. The real market fills you only if everyone ahead of you got done first — and it fills you mostly when you're wrong.

2026-08-07 · 6 min read
The best of 200 random strategies 200 variants tested — all of them pure noise what noise alone scores your "winner" · 2.0 expected best of 200 random tries = 2.1 0123 Sharpe ratio
Quant

The multiple-testing trap and the deflated Sharpe

A Sharpe of 2.0 means nothing until you say how many strategies you tried. Test enough coin flips and one of them looks like a genius — that's arithmetic, not alpha.

2026-08-06 · 6 min read
An overfit peak versus a robust plateau in parameter space overfit — a lonely peak lookback → every neighbour loses — it's noise robust — a broad plateau lookback → neighbours agree — sit in the middle
Quant

Parameter plateaus, not peaks

If your strategy makes money at a 14-day lookback and loses at 13 and 15, you haven't found a parameter. You've found a coincidence with a suspiciously specific address.

2026-08-05 · 5 min read
Confidence in a measured edge narrows with sample size measured edge · 95% confidence interval 0 true edge lower bound finally clears zero could still be worthless 301003001,0003,000 trades
Quant

How many trades is enough?

Thirty winning trades feels like proof. Statistically it's barely a rumour. Here's roughly how many you need before an edge can be told apart from a lucky streak.

2026-08-04 · 6 min read
The four rungs from backtest to full size 1 · Paper live data, no money 2 · Shadow signals, no orders 3 · Small live min size, real fills 4 · Full size scale slowly reconcile ✓ reconcile ✓ reconcile ✓ promote on tracking error — never on P&L
Quant

The deployment ladder — paper, shadow, small, scale

The gap between a good backtest and real money is not a switch you flip. It's four rungs, and each one exists to find the bug that would otherwise have cost you.

2026-08-03 · 6 min read
The order state machine, including crash recovery on restart: ask the broker PENDING SUBMITTED PARTIAL FILLED CANCELLED / REJECTED you died here the true state the broker is the source of truth — your database is a cache
Quant

From signal to order — an execution engine that survives a restart

A signal is an opinion. An order is a fact with your money attached. The code between them is the least glamorous and most dangerous part of any trading system.

2026-08-02 · 7 min read
Pre-trade checks and the breaker checks run in the order path — the strategy cannot skip them new order ✓ max size ✓ price sanity ✓ submit rate ✗ daily loss breaker trips flatten + halt one failed check stops everything — no override, no retry
Quant

Risk limits and the kill switch

Every algo eventually does something insane. The only question is whether a hard limit catches it in one second, or your broker calls you in an hour.

2026-08-01 · 6 min read
Live equity breaching the Monte Carlo floor Monte Carlo expected range 95th pct floor live equity breach → halt and review time →
Quant

Strategy decay — knowing when to turn it off

Every edge dies. The hard part isn't accepting that — it's telling a normal losing streak from a corpse, ideally before you've handed back a year of profit.

2026-07-31 · 6 min read
Temperature reshapes the token distribution prompt: "the sky is ___" low temp · sharp bluegreycleardark high temp · flat bluegreycleardark
AI & LLMs

Temperature and sampling — how an LLM picks its words

A model doesn't output a word — it outputs a probability for every possible next word. Temperature is the dial that decides how boldly it gambles when it turns those odds into an actual choice.

2026-07-30 · 5 min read
The agent loop: reason, act, observe, then answer Reason Act call a tool Observe tool result 1 · decide 2 · run 3 · feed back Answer when confident
AI & LLMs

How AI agents work — the reason–act loop

An "AI agent" isn't a smarter model — it's the same model put in a loop and handed tools. The whole trick is letting it think, act, see the result, and decide what to do next.

2026-07-29 · 6 min read
A random walk that looks like a trend the "trend" your eye draws… …is just accumulated coin flips
Quant

Why prices look random — the random walk

Stare at a price chart and you'll see trends, support, momentum — patterns everywhere. A lot of what you see is your brain inventing structure in what is, statistically, very close to a coin flip.

2026-07-28 · 6 min read
Expected value from a 40 percent win rate 0 +12 wins 40% × 30 −9 losses 60% × 15 +3 net EV per trade
Quant

Expected value vs win rate — why a 40% strategy wins

"I win 70% of my trades" sounds like a great system and tells you almost nothing. Win rate is the number beginners chase; expected value is the one that actually pays.

2026-07-27 · 5 min read
A lost update without isolation T1 (add 20) T2 (subtract 10) time read 100 read 100 write 120 write 90 final 90 (should be 110) T1's +20 lost
Backend & Data

ACID transactions and isolation levels

"The transaction went through" quietly promises four things. The trickiest is isolation — and loosening it is how two perfectly correct pieces of code combine to silently lose your data.

2026-07-26 · 6 min read
Token bucket versus leaky bucket Token bucket tokens added @ fixed rate request spends a token bursts OK until empty Leaky bucket requests pour in overflow dropped leaks @ fixed rate
Backend & Data

Token bucket vs leaky bucket — how rate limiters work

"100 requests per minute" sounds simple until you ask: 100 evenly spaced, or a burst of 100 at once? The two classic rate-limiter algorithms answer that question very differently.

2026-07-25 · 5 min read
Sharding rows across three machines by a hash of the key router hash(user_id) shard A users 0–3M shard B users 3–6M shard C users 6–9M each row lives on exactly one shard
Backend & Data

Database sharding — splitting data across machines

One database eventually hits a wall — too much data, too many writes for a single machine. Sharding is how you spread it across many, and the hard part is choosing where each row lives.

2026-07-24 · 6 min read
Adding and checking an item in a bloom filter add "apple" 0 1 0 1 0 0 1 0 h1 h2 h3 check "grape" 1 ✓ 0 ✗ 1 ✓ one bit is 0 → "grape" is definitely NOT in the set
Backend & Data

Bloom filters — saying "probably" to save a lookup

A tiny structure that answers "have I seen this before?" using a few kilobytes instead of gigabytes — by being willing to occasionally say yes when the answer is no.

2026-07-23 · 5 min read