AlphaFeed by Saxx

Every alert AlphaFeed sent

AlphaFeed sent an alert when a new Solana memecoin looked good. Each dot is one of its 60,724 alerts. The line is the price when the alert went out. The dot is the best price the coin reached after.

Scroll to see a day later

Most went nowhere. In AlphaFeed's own checks, 69% never rose above the alert price. They make the bright line.

About one in ten doubled at some point. A few went much further. mexicanunc went from a $7,465 market cap to $16 million.

0%below the alert price

But you only make money when you sell. Here is each coin's price one day after its alert.

The typical alert was down 23% after a day. One in seven had lost 90% or more.

The big winners were real. mexicanunc was still up 1,422x a day later. There were just too few of them to pay for the rest.

Six ways to sell, six losses

I wanted to know if any exit rule could beat this. So I replayed 1,893 alerts that passed a stricter filter, with $10 per trade and 4% fees and slippage on every buy and every sell.

Result after fees for each exit rule, about 1,700 trades each
  1. Trailing stop with a break-even floorfollows the peak, never sells below the entry after 1.5x−$1,327
  2. Sell everything at 3x−$2,440
  3. Sell in three stepshalf at 1.5x, 30% at 3x, the rest follows the peak−$2,617
  4. Sell everything at 2x−$2,941
  5. Sell everything at 1.5x−$3,449
  6. Trailing stopsells 40% below the highest price once it passes 1.5x−$5,548

Every rule also had a stop-loss at −30%. For all six, the typical trade ended at that stop-loss.

The result that almost fooled me

One test without a stop-loss made money:

First result: +$2,510counting 407 trades
Corrected result: −$7,675counting all 1,889 trades

It only counted the 407 trades that reached a sell target. The other 1,486 never did. With no stop-loss they stayed open while the coin died, and the test skipped them. Closed at their last price, the same strategy lost $7,675.

What AlphaFeed was

A Telegram bot I built and ran alone. It listened to pump.fun for new Solana tokens, checked each one for rug risks, scored it with a machine learning model and sent the best ones to subscribers. It checked 3.6 million tokens in six months. 26 people used the free beta. Nobody paid, and after these tests I did not want anyone to.

Python, FastAPI, PostgreSQL, Redis, XGBoost and LightGBM, and a Telegram bot, all on one small server with 4 GB of memory.

The AlphaFeed landing page headline: The AI that hunts alpha while you sleep, with buttons to apply for the free beta and to preview the bot on Telegram.
The landing page while it was live.