Kronos against 500 random stock-pickers

Kronos is an open-source foundation model for candlestick data. I tested it on Indian stocks in four rounds: do its forecasts rank stocks better than simple baselines, and does that hold up after fees and tax? I wrote the pass/fail criteria down before the first result. Nothing passed.

One result does look like a win. In the last round I simulated a ₹10 lakh portfolio holding the ten NIFTY 50 stocks Kronos ranked highest, rebalanced weekly, with Indian retail costs and tax. It returned +9.2% annualised net, against +5.6% for an index ETF. After tax it was still ahead by 1.9 percentage points.

An equal-weight basket can beat a cap-weighted index with no skill at all, though, so I also graded the book against 500 random stock-pickers. Of those 500, 79 did as well or better. The book came in at the 84th percentile.

Ranking 50 names didn’t help choose the ten. Kronos mostly sorted the losers to the bottom, and at the top, the only part a ten-stock book buys, it added nothing (+3 bps spread, t = 0.53). The signal was faint anyway. Read from the model’s hidden state, its rank IC on the NIFTY 50 (how well predicted and actual rankings agree) was +0.022, and 50 names is too few to detect a signal that size.

A futures version failed too: a 4 bps gross edge against 432 bps of weekly standard deviation, and one week at −18%.

None of it is live trading. The data covers August 2023 to August 2026, a strong bull market for Indian mid and large caps.

The earlier rounds are in the case study, including the finding that about 70% of the model’s own score was reversion toward its 90-bar normalisation window.