Indonesia stock prediction (IDX): the honest numbers and why my model lands below even

I run Trading Agent, and I publish every prediction my model makes — including the bad ones — at /predictions. Most founders writing about their own product quietly pick the markets where they look good. This is not one of those articles. Indonesia is one of the markets where my model does worse than a coin flip, and I think the honest version of that story is far more useful than a sales pitch.

The number, plainly

On the Indonesia Stock Exchange (IDX) — the .JK tickers — my model has now logged roughly 1,950 verified directional predictions with an accuracy of about 46.3%. That is not a small, early sample anymore. It is a large, settled one, and what it has settled at is below even. Not near a coin flip — under it. On a base this size, a reading that far below 50% is real: in Indonesia, my model is, on the raw directional call, worse than chance for now.

For context, my blended accuracy across all 16 markets is about 49.7% — already barely better than a coin toss — and my two strongest markets, Canada and the US, sit around 53%. At the bottom sits Thailand at about 43.8%, and Indonesia is among the weakest too, sitting clearly below the blended average. I'm not going to dress that up. The model has nearly two thousand verified calls here and it lands below even.

The interesting question isn't whether it's below a coin flip — nearly two thousand verified predictions show that it is. The interesting question is why. And the answer is almost entirely about market structure.

Why a momentum model lands below even in Indonesia

My model, like most machine-learning systems applied to equities, is fundamentally a pattern-and-momentum engine. It looks for trends, continuation, and the statistical fingerprints of price moving in a direction and keeping moving. That approach works best in deep, liquid markets full of momentum-driven single stocks where a trend, once started, tends to persist. Indonesia is an emerging market, and several of its structural features run directly against what the model is built to read.

It's a bank- and commodity-heavy index. A large share of IDX's weight sits in a handful of big banks and commodity-linked names — the major lenders and a few resource and telecom heavyweights being the obvious examples. Bank prices in an emerging market move on rate expectations, credit cycles, and policy more than on any clean chart pattern. Commodity-linked names move on global coal, nickel, and palm-oil prices set far away from Jakarta. In both cases the real driver is exogenous — something the model can't see on the local price series — so the technical signal it thinks it has is being overwritten by forces off the chart.

Momentum persistence is lower. In deep momentum markets, a move that starts tends to continue for long enough that a model can catch it. In Indonesia, single-stock trends are choppier and reverse faster. The clean, repeated directional continuation my model is built to detect simply happens less often, so more of its calls are made on patterns that don't follow through.

Liquidity is thin below the top names. The largest few stocks trade actively, but liquidity drops off sharply below them. Thin order books mean prices gap, jump on small flows, and move on liquidity events rather than on the gradual, trend-forming behaviour a momentum model needs. A stock that lurches because one institution rebalanced isn't trending — it's just moving — and the model can mistake that noise for signal.

The rupiah and foreign flows inject noise the model misreads. This is the big one. Indonesian equities are unusually sensitive to the rupiah and to foreign portfolio flows. When the currency weakens or global risk appetite turns, foreign investors pull money out across the board, and prices move together on a macro and FX impulse that has nothing to do with any individual company's chart. To a momentum model reading local price patterns, that looks like a directional move it should be able to extrapolate — but it's currency-and-flow noise, and extrapolating it is exactly how you end up below even. The model misreads an FX-driven, correlated sell-off as a stock-specific trend.

So when I mention a name like one of the big banks or telecoms as an example, I'm using it as a structural example of these dynamics — never as a pick. The point is the category behaviour, not the ticker.

What I actually do with a market that's below even

I label Indonesia signals Bearish, Neutral, or Bullish — never Buy or Sell, never a guarantee, and never a "90% confident" number — and on IDX I treat them with heavy skepticism, because the verified data tells me to. And I want to be careful here: a sub-50% market is not a free contrarian edge you can just invert. Once you account for the costs and the noise, "below even on the raw call" doesn't reliably flip into "above even by doing the opposite" — it mostly tells you the technical signal here is weak and contaminated by forces the model can't observe. The honest conclusion is narrower and less exciting than any marketing line: in an emerging market dominated by banks and commodities, with thin liquidity and heavy currency and foreign-flow sensitivity, my model is below a coin flip for now, and I tell you that rather than hiding it behind a blended average.

One more thing worth being explicit about: the numbers above are the live record. Backtests on this kind of system always look better than reality — it's easy to make a curve fit the past. The live, verified directional accuracy is the real number, and it's the lower one. The ~46.3% I'm quoting for Indonesia is what actually happened on predictions I made before I knew the outcome, not what a polished backtest suggests. A backtest of this market would look kinder; the live record is the one I stand behind, and it's below even.

This is the whole reason I built the brand around radical honesty. Plenty of AI stock tools quietly bury their losers — I wrote about that pattern in why most AI stock-picking tools are lying. I'd rather show you a below-even market and explain the structure behind it than show you a polished number I can't stand behind. If you want the full breakdown of how the model is built, scored, and verified, it's all on the methodology page.

Indonesia is, for now, a market where my model's most valuable contribution is honesty about where it stands: below even, for structural reasons, and said out loud. That's not the answer a marketer wants. It's the answer the data supports.

---

This article is educational content about machine learning and market structure. It is not financial advice, not a recommendation to buy or sell any Indonesia-listed or other security, and not directed at any individual's circumstances. Trading Agent is a quantitative research tool operated by WU Capital Limited (New Zealand).